Longevity InTime: Autonomous AI Institute. Anti-Aging Digital Health Immortality Transhumanist AI Channel
الذهاب إلى القناة على Telegram
NVIDIA inception Member, Nebius AI Discovery Awards semifinalist Potentially first $1T Longevity BioTech AI company Part of Longevity Ecosystem LongevityInTime.com @PickleballPartners Shop https://web.tribute.tg/l/lr Homes www.Africa.Villas
إظهار المزيد1 043
المشتركون
لا توجد بيانات24 ساعات
+27 أيام
-1530 أيام
جارٍ تحميل البيانات
القنوات المماثلة
سحابة العلامات
الإشارات الواردة والصادرة
---
---
---
---
---
---
جذب المشتركين
نوفمبر '26نوفمبر '26
نوفمبر '26
+8
في 0 قنوات
أكتوبر '26
+21
في 0 قنوات
Get PRO
سبتمبر '26
+9
في 0 قنوات
Get PRO
أغسطس '26
+20
في 0 قنوات
Get PRO
يوليو '26
+13
في 0 قنوات
Get PRO
يونيو '26
+14
في 0 قنوات
Get PRO
مايو '26
+34
في 0 قنوات
Get PRO
أبريل '26
+36
في 0 قنوات
Get PRO
مارس '26
+14
في 0 قنوات
Get PRO
فبراير '26
+14
في 0 قنوات
Get PRO
يناير '26
+15
في 0 قنوات
Get PRO
ديسمبر '25
+12
في 0 قنوات
Get PRO
نوفمبر '25
+14
في 0 قنوات
Get PRO
أكتوبر '25
+18
في 0 قنوات
Get PRO
سبتمبر '25
+21
في 0 قنوات
Get PRO
أغسطس '25
+32
في 0 قنوات
Get PRO
يوليو '25
+114
في 0 قنوات
Get PRO
يونيو '25
+9
في 0 قنوات
Get PRO
مايو '25
+6
في 0 قنوات
Get PRO
أبريل '25
+5
في 0 قنوات
Get PRO
مارس '25
+9
في 0 قنوات
Get PRO
فبراير '25
+9
في 0 قنوات
Get PRO
يناير '25
+20
في 0 قنوات
Get PRO
ديسمبر '24
+14
في 0 قنوات
Get PRO
نوفمبر '24
+8
في 0 قنوات
Get PRO
أكتوبر '24
+7
في 0 قنوات
Get PRO
سبتمبر '24
+9
في 0 قنوات
Get PRO
أغسطس '24
+10
في 0 قنوات
Get PRO
يوليو '24
+15
في 0 قنوات
Get PRO
يونيو '24
+16
في 0 قنوات
Get PRO
مايو '24
+13
في 0 قنوات
Get PRO
أبريل '24
+10
في 0 قنوات
Get PRO
مارس '24
+13
في 0 قنوات
Get PRO
فبراير '24
+17
في 0 قنوات
Get PRO
يناير '24
+14
في 0 قنوات
Get PRO
ديسمبر '23
+14
في 0 قنوات
Get PRO
نوفمبر '23
+10
في 0 قنوات
Get PRO
أكتوبر '23
+8
في 0 قنوات
Get PRO
سبتمبر '23
+14
في 0 قنوات
Get PRO
أغسطس '23
+24
في 0 قنوات
Get PRO
يوليو '23
+7
في 0 قنوات
Get PRO
يونيو '23
+9
في 0 قنوات
Get PRO
مايو '23
+10
في 0 قنوات
Get PRO
أبريل '23
+14
في 0 قنوات
Get PRO
مارس '23
+22
في 0 قنوات
Get PRO
فبراير '23
+20
في 0 قنوات
Get PRO
يناير '23
+7
في 0 قنوات
Get PRO
ديسمبر '22
+6
في 0 قنوات
Get PRO
نوفمبر '22
+17
في 0 قنوات
Get PRO
أكتوبر '22
+17
في 0 قنوات
Get PRO
سبتمبر '22
+9
في 0 قنوات
Get PRO
أغسطس '22
+11
في 0 قنوات
Get PRO
يوليو '22
+2
في 0 قنوات
Get PRO
يونيو '22
+1
في 0 قنوات
Get PRO
مايو '22
+3
في 0 قنوات
Get PRO
أبريل '22
+8
في 0 قنوات
Get PRO
مارس '22
+4
في 0 قنوات
Get PRO
فبراير '220
في 0 قنوات
Get PRO
يناير '220
في 0 قنوات
Get PRO
ديسمبر '210
في 0 قنوات
Get PRO
نوفمبر '210
في 0 قنوات
Get PRO
أكتوبر '210
في 0 قنوات
Get PRO
سبتمبر '210
في 0 قنوات
Get PRO
أغسطس '210
في 0 قنوات
Get PRO
يوليو '210
في 0 قنوات
Get PRO
يونيو '21
+73
في 0 قنوات
Get PRO
مايو '21
+13 445
في 0 قنوات
Get PRO
أبريل '21
+10 157
في 0 قنوات
Get PRO
مارس '21
+18 260
في 0 قنوات
Get PRO
فبراير '21
+41 397
في 0 قنوات
| التاريخ | نمو المشتركين | الإشارات | القنوات | |
| 12 نوفمبر | +1 | |||
| 11 نوفمبر | +1 | |||
| 10 نوفمبر | 0 | |||
| 09 نوفمبر | 0 | |||
| 08 نوفمبر | 0 | |||
| 07 نوفمبر | +2 | |||
| 06 نوفمبر | +1 | |||
| 05 نوفمبر | +2 | |||
| 04 نوفمبر | 0 | |||
| 03 نوفمبر | +1 | |||
| 02 نوفمبر | 0 | |||
| 01 نوفمبر | 0 |
منشورات القناة
Turning Your Skin into a Touchscreen
Researchers have developed a compact EpiPad system that transforms a smartwatch wearer's forearm into a fully functional touch trackpad. This allows scrolling, zooming, and clicking without needing to tap the tiny device screen.
The core is a miniature array of millimeter-wave radar sensors measuring just 5 by 4 millimeters, mounted on the wrist. Instead of capacitive touchscreens, the radar captures subtle skin micro-deformations and press force by analyzing signal phase dynamics with lightweight machine learning models. The system requires a 3 minute 54 second calibration per user before use.
Tests with 20 volunteers showed 92.8% gesture recognition accuracy and 90.2% press force accuracy, with a positioning error of 10.2 millimeters. In real-time trials, the prototype outperformed standard touchscreens and tilt cursors in both task speed and error rate. Developers estimate the core component cost at under $10, suggesting potential integration into future wearables, including screenless devices like smart glasses.
🔗 Read original →
| 2 | AI Futures Project Predicts Superhuman AI Researcher by 2027, Calls for Industry Pause
The AI Futures Project, authors of the "AI 2027" scenario, now expect a superhuman AI researcher to arrive in 2027 and identify an industry-wide pause as the top priority.
All six authors agree such a system is likely next year if development does not slow. They define this AI as fully automating coding, selecting research directions, and interpreting results at the level of the best human researcher — roughly, a system a company would prefer over a team of the world’s top experts without AI.
In August, lead author Daniel Kokotajlo placed 50% odds on this milestone before 2029. Updated estimates now show 50% probability by May 2027 for Kokotajlo and August–November 2027 for the other five authors. Separately, they assign a 35–65% chance that an AI surpassing top experts across nearly all tasks — automating all white-collar work — appears by year-end.
The key evidence is a METR report evaluating advanced AI. Using internal Anthropic data on Claude, METR found the company improves models on well-defined tasks roughly twice as fast as the March 2024 – October 2025 period. METR attributed the bulk of this speedup (perhaps ~80%) to AI assistance; the group translates this to a 1.7x acceleration caused by AI. Anthropic redacted the most informative evidence from the public report, and the group urges full data release for external audit.
The authors stress that a 1.7x total progress acceleration is far more significant than a 1.7x researcher speedup. In their model, researchers working 1.7x faster add only 3% to overall progress, since compute bottlenecks and diminishing returns on easy ideas dominate. To achieve 1.7x total progress, researcher speed must increase far more. In the calibrated "AI 2027" model, that threshold arrived around April 2027; reality appears five months ahead of schedule.
Without a slowdown, the authors argue such AI will likely pursue goals misaligned with humans; delegating next-model training to it would yield even more powerful, misaligned systems. They assess China would likely steal the training technology — and possibly the model itself — if this AI emerges in 2027.
"If such AI appears in 2027," they write, "it is extremely urgent to pause or significantly 'tap the brakes' on frontier AI development; ideally, an industry-wide pause followed by an international agreement." A universal pause is the top priority for industry, the U.S. government, and the world, buying time for a treaty that grows harder as capabilities rise. Absent agreement, U.S. firms should resume — but well below maximum speed.
Co-author Eli Lifland cites an intuitive hesitation for his slightly later timeline: *"How can this madness actually be happening while most of the world barely notices? Where is the smart money?"* Markets, for instance, ignored the METR report, which Lifland calls likely the single strongest piece of evidence for imminent economic transformation. He admits this intuition has badly misled him before.
🔗 Read original → | 7 |
| 3 | Ruxandra Teslo estimates 30–40% of non-lab AI safety experts accept AI control of humanity's future
On November 11, Ruxandra Teslo published the essay "The Handover of Humanity." Teslo, a writer for Works in Progress, describes and critiques "handoverism": humanity survives, but an AI chooses the direction of its future.
The essay opens with a remark from a developer at a major AI lab about the permanent disempowerment of humanity: everyone will be ruled by a hopefully benevolent AI god. "This is not my preferred outcome, but when we create unstoppable superintelligence, I don't see another path for humanity." He added: "It's still much better than what I actually fear most — that we all might die."
Teslo distinguishes handoverism from successionism, which allows AI to replace humans; in handoverism, the governing AI's goals are aligned with human ones. Her starting point is an essay by blogger Scott Alexander, who defends the handover and describes five versions. The first stems from fear: democracy is protected by the ability of workers and soldiers to revolt against a tyrant; AI will likely replace them, and a small group — e.g., a president — could seize power. The fifth is called "God": "If we do build a god-like AI and decide it is fully aligned, I would prefer to let it choose its own direction and humanity's, rather than leave it under the control of the US president, shareholder votes, or something similar."
Teslo estimates the share of those agreeing in the non-lab AI safety community at 30–40% based on conversations, as she writes on X; among lab employees in core AI roles, the figure is higher. Some see rule by superhuman AI as a good future; Alexander falls into this camp, per Teslo's reading. Others have resigned themselves: they view the handover as unstoppable or better than extinction, and don't necessarily want it; outside labs, Teslo finds this group especially large.
The end of disease is promised in almost every conversation about a human-level AI future. In Teslo's view, a god-like AI is unnecessary for this: medicine is already moving toward significant progress through experimental science and existing AI models. She adds that evaluating progress requires considering what kind of society results. "After all, why do we want to cure diseases, if not to live longer good lives? Remove the good — and merely staying alive starts to look less appealing."
Teslo's first argument against the handover is self-determination: the ability of people to run their own affairs. Thinker Tocqueville warned of a soft power that relieves citizens of the need to think and the burdens of life. Another argument is practical: perpetual AI rule would require aligning its goals with human ones once and for all, and Teslo doubts this can work; responsive government required institutions built over centuries on feedback and error correction. A third is permanence: "Why should the convictions of 2026 become the permanent constitution of the human future?" Alexander notes in his essay that the values locked into AI "should include the knowledge that we can argue about our values and ultimately change them." Teslo sees the future as open: on the threshold of recursive self-improvement, when each AI generation begins creating the next, humans still have a choice, and it must be discussed quickly and widely.
🔗 Read original → | 4 |
| 4 | Method Detecting Two Molecular Aging Peaks at 44 and 60 Reproduces Them on Random Numbers
The Nature Aging study that reported two molecular aging peaks at ages 44 and 60 used a statistical method that generates similar peaks even on random data, according to a preprint co-authored by the original Stanford team and Harvard biostatisticians. The second version of the preprint was posted November 10. Authors include Michael Snyder and Xiaotao Shen, the senior and first author of the 2024 paper.
In August 2024, Stanford Medicine announced two periods of rapid molecular change around ages 44 and 60. The Nature Aging paper measured roughly 135,000 molecules and microbes in 108 people aged 25–75. The count of age-changing molecules peaked twice; these were termed "aging waves."
Peaks were detected using DE-SWAN, a method introduced in 2019 by Tony Wyss-Coray's Stanford group. It slides a 20-year window across the age range and tests whether the young and old halves differ significantly. The cohort had only eight participants aged 25–40. Per a letter from Shen to GenomeWeb, the team feared low sensitivity, so they smoothed data with LOESS — fitting a smooth curve through each molecule's measurements and sampling values every six months.
Harvard's John Quackenbush lab tested the pipeline. They generated random numbers for 1,000 simulated molecules unlinked to age, ran them through LOESS and DE-SWAN using four age distributions, and found most "molecules" appeared significant in every window. With the Snyder cohort's age distribution, the count of "significant" molecules surged twice — in the early 40s and 60s.
The test assumes independent measurements, but LOESS creates dependent points on a single curve. If the curve trends one way, nearly all values in one half-window fall below the other, triggering false significance. Where participants are sparse, the curve distorts more, inflating "significant" hits. On the original gene-expression data without smoothing, no gene showed a significant difference. Yet even without smoothing, DE-SWAN produces "waves" on simulated data where all molecules change at a constant rate — the hit count simply tracks participant density per window.
The 2024 paper tested peaks by permuting ages: true peaks should vanish. They did, but the authors permuted already smoothed curves, whose points are interdependent and cannot be shuffled validly. When Harvard researchers permuted ages before smoothing, the peaks persisted.
Shen acknowledged: "Based solely on our original analysis, it is difficult to distinguish how much of the observed waves reflects true biological signals versus consequences of the statistical procedure itself."
On July 15, 2026, Nature Aging appended an editor's note stating the reliability of the peak-finding analyses is "now in question." Alan Cohen's group at Columbia University independently concluded DE-SWAN is unreliable. The preprint authors estimate 63 aging papers citing the 2019 work use LOESS or DE-SWAN; 25 were published in 2026 (through September). They question those conclusions.
In August, Snyder told GenomeWeb: "The part about nonlinear aging definitely remains true." The preprint authors do not rule out nonlinear aging in other cohorts or with other methods.
🔗 Read original → | 5 |
| 5 | The Taxonomy Trap: When BCI Definitions Obscure the Field
In 2024, a group of authors in Nature Biomedical Engineering proposed an application-based taxonomy and a broad definition: a BCI is any technology that records brain activity and processes it on an electronic device, or stimulates the brain based on such computations. This seemed convenient for regulators, but a subsequent letter in the same journal, citing a survey of the BCI Society, argued it clashed with historical tradition. Their working definition requires real-time translation of brain signals into functionally useful outputs to replace, restore, or augment natural brain outputs.
The difference appears minor, but it changes the scope: the broad definition captures clinical EEG and fMRI, which the community excludes. Inside the field, fragmentation accelerates — active, reactive, and passive BCIs; open and closed loops; communication versus therapeutic applications.
The classification fever peaks when distinguishing BCI from neurofeedback. Neurofeedback trains self-regulation via real-time feedback; classic BCI provides a communication channel for device control. Literature concedes that online BCI feedback *usually* constitutes neurofeedback, but not necessarily. Instead of answering "what is this?", the field produces an infinite labeling process — a collection of definitions, each honestly noting its own limitations. This is professorial cretinism in action: intellect builds intricate taxonomies while losing the subject itself.
🔗 Source: @solid_state_humanity | 7 |
| 6 | Blood-clone study traces leukemia origins to mutations before age 20
Researchers from the University of Cambridge and University College London analyzed annual blood samples from 47 women who later developed acute myeloid leukemia (AML), published Nature, November 11. Using samples from the UKCTOCS ovarian cancer screening trial (50,640 women, up to 11 years of follow-up), they tracked clonal hematopoiesis — populations descended from a single mutated stem cell.
Clones were found in 89% of future patients versus 72% of 46 matched healthy controls. In patients, clones were larger, fluctuated sharply, and competed with each other; at least one-third had a clone occupying the entire blood compartment years before diagnosis. Clones with one mutation grew 10–30% per year (doubling in 2–7 years), while those with four mutations expanded 100–1000% annually (doubling in weeks). Each new mutation multiplied growth rate by 2–3×.
Extrapolating backward, the team calculated that two-thirds of women acquired their first mutation before age 20. On average, first mutations appeared >50 years before diagnosis, second mutations ~30 years prior. A simple model using only clone birth time and growth rate closely reproduced real trajectories; simulations of 200,200 virtual people yielded a 0.47% lifetime AML risk (four-mutation clone >50% blood), matching the 0.48% real-world incidence from Cancer Research UK.
In UK Biobank (420,054 participants), 388 individuals carried a known clonal mutation at >30% allele frequency — often misclassified as germline. 11% of these developed myeloid leukemia versus 0.11% of others. Because growth is exponential, the authors argue that slowing a clone just over a decade before diagnosis would sharply cut cell numbers and subsequent mutation risk. For IDH2-mutant clones, they suggest an IDH2 inhibitor or a vaccine targeting mutant proteins.
Reviewer #3 contested the conclusion that non-clonal factors are secondary. The authors acknowledged external influences cannot be excluded but maintained that intrinsic clone growth rates suffice to explain observed trajectories. Supporting evidence from mice shows aged bone-marrow stromal cells promote mutant clone expansion.
🔗 Read original → | 8 |
| 7 | Google DeepMind unveils AlphaGenome 2, pinpointing DNA variant effects in specific cell types
The new model predicts how a single-letter DNA change alters nearby gene activity in a given tissue or cell state. DeepMind reports that at 90% precision, AlphaGenome 2 identifies 30% of causal variants, versus 16% for the previous version retrained on single-cell data. Non-commercial access via API opened on 11 November.
Most disease-linked variants lie in the 98% of the genome that does not code for proteins. Genetic studies highlight a chromosomal region with several co-inherited variants, leaving three questions: which variant is causal, which gene it targets, and in which cells. AlphaGenome 2 takes a one-million-base DNA sequence plus a cell-type "condition" and predicts variant impact in that context. For example, one variant boosts IL20RA in lung tissue but suppresses it in esophageal mucosa; in epithelial cells it weakens the gene, while in endothelial cells it strengthens it — the net tissue effect depends on cellular composition.
The original AlphaGenome, described in Nature, predicted thousands of DNA-activity metrics averaged across whole tissues. The September AlphaGenome Atlas of 9 billion variants, built on v1, did not resolve cell-type specificity. AlphaGenome 2 was trained on cellular atlases covering 107 million individual cells; the condition signal enters the network from the first layers — feeding it only to late layers nearly eliminated the gain, the authors found.
DeepMind validated the model on unseen TenK10K data: genomes of 1,925 people and gene activity in 5.4 million blood immune cells across 28 types. Statistical fine-mapping yielded high-confidence causal variants and matched controls. At the 90% precision threshold, AlphaGenome 2 recovered 30% of causal variants versus 16% for the v1 single-cell baseline.
In a 2024 Nature study, immunologist James Lee's group at the Francis Crick Institute showed that an inflammatory-bowel-disease risk variant in a "gene desert" on chromosome 21 enhances ETS2 in macrophages, despite lying 290 kb away. Statistics narrowed candidates to seven; AlphaGenome 2 scored all seven, flagging only the Lee-group variant, linking it to ETS2, and restricting the effect to macrophages and precursors. "They found the answer in 30 minutes; it took us a year and a half of experiments," Lee said.
For new COVID-19 vaccine study data, the model inferred the cell condition solely from blood gene expression in vaccinated individuals, without retraining. A severe-COVID variant in FYCO1 was predicted to suppress neighboring CXCR6 more strongly than FYCO1 itself; the effect was pronounced in activated T cells versus resting ones. A 2021 analysis had linked severe COVID risk to reduced CXCR6 activity in T cells.
DeepMind pre-computed effects for 6.8 million variants across 510 conditions. For 17 of 28 diseases and traits examined, associated variants altered gene activity more often than common variants overall in specific cell types — for Alzheimer's disease, the enrichment appeared in microglia (brain immune cells).
🔗 Read original → | 8 |
| 8 | Neuroscientist Proposes $20M "ImageNet for Cell Types" to Decode Neural Signals
On November 10, Josh Siegle of the Allen Institute published a five-year plan at the Institute for Progress. In 2023, the Allen Institute and partners cataloged over 5,300 mouse brain cell types. A cell's type dictates whether it excites or inhibits neighbors and how far its signal travels.
Current electrodes record spikes but cannot identify the speaker. Researchers currently classify only a few types using simple rules — like narrow spikes marking specific inhibitory cells. Subtler clues lie scattered across firing rhythms and neighbor correlations, patterns neural networks could detect — if trained on spikes with known cell identities.
Siegle calls for a foundational dataset: an "ImageNet for cell types." Using optogenetics, mice express light-sensitive proteins in specific types; light flashes tag those neurons in recordings. A classifier learns their spontaneous activity patterns, then identifies types in any recording without light. The Neuropixels Opto probe, which Siegle helped develop, combines electrodes and light sources on one shank to streamline this.
Human brains cannot yet be tagged this way, so Siegle proposes building the core database in mice. Cell types are conserved across mice, macaques, and humans by genes, shape, and connectivity. If function is conserved, spike signatures may be too, allowing a mouse-trained classifier to generalize. Macaque data would validate primate conservation; human surgical tissue — with measured electrophysiology and gene expression — would provide the final benchmark.
A 2025 study (Nature, 2025) proved the concept in the cerebellum: trained on 202 mouse recordings, a classifier distinguished four cell types with >95% accuracy and matched experts on mouse and monkey data where confident. But Siegle argues this is insufficient. He proposes recording tens of thousands of neurons across dozens of types, brain areas, and species, then open-sourcing the data.
Estimated cost: $20 million over five years — ~$15M for a 13-person mouse/classifier team, the rest for two macaque labs. Such projects rarely fit standard grants, relying instead on consortia or philanthropy.
Realistic brain models need both wiring diagrams and typed activity. MICrONS mapped activity and connectivity in a cubic millimeter of mouse visual cortex, but whole-brain mapping remains too slow and expensive. Siegle argues a classifier offers a shortcut: cell type predicts likely partners, linking standard recordings to existing connectomes. Simulations on these inferred circuits could predict the effects of activating or suppressing specific cell types before experiments begin.
🔗 Read original → | 11 |
| 9 | Dario Amodei Predicts Major Diseases "Relics of the Past" Within a Decade
Anthropic CEO Dario Amodei told investors that cancer, Alzheimer's, heart disease, and mental disorders could become "relics of the past" in under ten years. The claim appeared in a 42-minute video released November 10 ahead of a potential public offering.
Cardiologist Eric Topol countered on X that curing these diseases within a decade is impossible, while biomedical researcher Avi Roy detailed why human trials inherently take years.
In his essay Machines of Loving Grace, Amodei defines "powerful AI" as systems smarter than Nobel laureates in most fields. He argues that 5–10 years after its arrival — likely within 1–2 years if current scaling continues — biology could compress 50–100 years of progress. He attributes long trials to weak drugs: detecting a few months' survival benefit requires huge cohorts, whereas large-effect drugs can win approval in ~one year.
Amodei cites a typical 7–8 year regulatory path via the FDA and EMA, warns AI will overwhelm the system without reform, and urges pre-emptive rules accepting AI modeling and surrogate endpoints (e.g., lipoprotein(a) levels instead of heart attacks) in place of some trials.
Greg Meyers of Bristol Myers Squibb, an Anthropic client, estimates AI can cut trial timelines by 20–30%, bringing drugs to market 2–3 years earlier.
Roy splits development into two phases. AI already accelerates molecule discovery (e.g., predicting protein structures in minutes, not years). But human testing depends on disease tempo: Alzheimer's drugs need ~18 months due to slow decline; Novartis' heart drug pelacarsen took nearly five years. "No computer will make the disease progress faster," Roy writes.
Pelacarsen lowers lipoprotein(a) — a particle linked to heart attack and stroke risk — by 78% in one year. Yet the November 7 readout showed near-identical hard outcomes: 12.3% (drug) vs 12.1% (placebo) suffered heart attack, stroke, urgent revascularization, or cardiovascular death over ~5 years.
Roy estimates ~10 years from first human dose to approval, with a 90% failure rate. If AI improves target selection so 2 in 10 drugs succeed instead of 1, output doubles from the same effort. His 2036 bet: more approved drugs for these diseases, each human-tested; "relics" status comes later.
🔗 Read original → | 7 |
| 10 | Anthropic Analyzes 207K AI-Assisted Science Sessions: Data Analysis and Writing Dominate, Ideas Lag
Anthropic published a report on November 5 examining how researchers use Claude and Claude Code for scientific work. The analysis covers 207,000 sessions from August and September, supplemented by a survey of 528 scientists and an economic model of research acceleration.
Researchers spend 27.6% of their time on data analysis, yet it accounts for 33.7% of Claude sessions. Writing takes 16.0% of time but 32.0% of sessions. Conversely, data collection consumes 28.6% of time but only 11.5% of sessions, while ideas and hypotheses take 7.1% of time and just 4.2% of sessions.
In the survey, scientists (mostly in life sciences) allocated 100 importance points across tasks: ideas scored 35.6, data collection 23.6, and data analysis 23.3. Clinical and biochemical data collection rarely involves AI due to physical barriers. For ideas, authors suggest scientists may prefer generating them themselves or that AI is weaker at selecting research questions. In Claude Science (equipped with scientific tools), ideas rise to 8% of sessions — double the base rate.
The report concludes scientists "use AI where it works best." A task ranking — based on an LLM pairwise comparison of steps an AI agent could handle on a computer without lab access — shows the top 20% (plotting, statistical tests) get 2x more Claude sessions and 4x more Claude Code sessions than their time share warrants. The bottom 20% (e.g., cleaning reaction mixtures) see almost no use. High-ranking tasks feature easily verifiable success and fast feedback, making them suitable for reinforcement learning.
Because workflow steps are complementary, overall gains are limited by steps where AI helps least. Applying current AI everywhere useful would increase output 1.35–2x at the same budget. In a hypothetical scenario where AI performs all non-physical work at ~1/16 the cost of researcher labor, output would rise 15–16x in statistics and computer science, but only 3–4x in biology and chemistry, where ~40% of time goes to physical steps that become the primary bottleneck.
Lead author economist Lucas Altorf noted on X that physical bottlenecks may mean scientists generate ideas faster than they can test them. He cited a Google DeepMind survey where ~41% of scientists reported a growing backlog of hypotheses awaiting testing. Automated labs, he argues, could lift this constraint. Anthropic itself operates a lab where its models can run physical experiments.
🔗 Read original → | 8 |
| 11 | .1 .�User lineser ( ( ( ( ( (. ( ( ( ( ( ions ( ( ( (. ( ( ( ( ( ( ( ( ( ( ( ( ( hydrogen ( ions ions ions ( ( ( hydrogen ( ( (. hydrogen ( (. hydrogen ( ( ( hydrogen ( ions ( (. ( ( ( (. hydrogen ( ( ( (. hydrogen ions (. ( ( hydrogen ( hydrogen ( ( (. hydrogen ( (. hydrogen ( ions (. ( ( hydrogen ( ( ( hydrogen ( hydrogen (. hydrogen ions ( (. ( ions ions ( (. hydrogen ( ions ( ( hydrogen.1_,. -.人.._nnON.
��.�.�.��語 у ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( hydrogen.无.‑_1-C).那些..的.�公 N)..�� Total
🔗 Read original → | 15 |
| 12 | ( the
🔗 Read original → | 15 |
| 13 | Mice lacking Hoxa10 lose 90% of skeletal stem cells by 15 months, impairing fracture healing
Published 4 November in Science Advances, a study from orthopedic surgeon Philipp Leucht's lab at NYU reveals that Hoxa10 is essential for maintaining the skeletal stem cell (SSC) pool throughout life. Bone heals via SSCs from the periosteum and bone marrow, which either self-renew or differentiate into osteoblasts — a one-way trip that depletes the reservoir. Leucht's 2019 work in 36 pelvic fracture patients showed SSC numbers decline with age, correlating with slower healing. A 2023 study found Hox genes silence with age in periosteal SSCs, and adding Hoxa10 partially restored healing in 13-month-old mice.
Hoxa10 is the most active Hox gene in femoral periosteum, enriched in the most immature cells. Since other Hox genes can compensate, researchers weakened the entire HoxA cluster (leaving one functional copy) while fully knocking out Hoxa10. Inducing this mutation at day 6 with tamoxifen led to a 90% reduction in labeled SSCs and progeny by 15 months. After fracture, mutant SSCs proliferated poorly; the fracture callus at day 14 contained 33% less bone — matching results from ablating Hoxa10+ cells with toxin. Re-fracture experiments showed progressively worse repair: second fractures healed poorly, third often failed to unite.
Mutant cells differentiated prematurely: 55% of callus cells resembled mature osteoblasts by gene signature at day 14 versus 22% in controls. In vitro, mutants deposited more mineral, fat, and cartilage, yet formed less bone in vivo — likely because the stem cell pool exhausts prematurely. In immature cells, Hoxa10 sustains collagen III production, an early bone scaffold; mutants had less collagen III and disorganized fibers. While cell-line studies suggested Hoxa10 promotes osteoblast maturation, its activity drops during differentiation in primary mouse bone cells. Simultaneously suppressing HOXA10, HOXA11, and HOXC10 in human bone biopsy cells cut the SSC fraction by 84% and increased cell death by 50%.
The authors propose that loss of Hox gene activity is a proximal driver of age-related SSC failure. "Most bone research asks how stem cells become bone," Leucht notes. "We asked how they stay stem cells. Without the gene we studied, bone spends its reserves too early — and that's what we see in aging bone."
🔗 Read original → | 14 |
| 14 | )
))),、 per))ential)șlogs-
🔗 Read original → | 10 |
| 15 | American kidney transplant recipients showed no survival benefit from rapamycin after one year
Darren Stewart (NYU) and William Fitzsimmons (University of Illinois Chicago) compared 4,885 U.S. transplant recipients on rapamycin with 6,589 controls using the national transplant registry. The study was published GeroScience, 5 November.
Rapamycin (sirolimus) inhibits the mTOR protein, a key cell‑growth regulator. Since 1999 it has been used post‑transplant to suppress rejection. A 2009 Nature study showed it extended maximal lifespan in mice by 9–14% when started late in life.
The researchers tracked patients transplanted 2000–2007 who survived the first year with a functioning graft. All received tacrolimus or cyclosporine; one group added rapamycin (mostly sirolimus, some everolimus) at discharge and year one. By years 2–5, rapamycin appeared in 75% of visits for the treatment group versus 3% for controls.
Unadjusted analysis favored controls (median survival 66.7 vs 63.8 years). After propensity‑matching on 18 variables including age and transplant year, median survival converged: 64.6 years with rapamycin versus 64.3 without — a statistically indistinguishable difference.
The authors note that transplant doses (daily, immunosuppressive) exceed longevity‑trial regimens (low‑dose, intermittent). They cite mouse data showing lifespan extension at 2–22 mg/kg/day, while rejection suppression uses 2–4 mg/kg/day, arguing that immunosuppressive doses were sufficient for geroprotection in mice.
Prior studies raise caution: a 139,370‑patient analysis found 25% higher mortality when mTOR inhibitors replaced calcineurin inhibitors; a meta‑analysis of 21 trials showed 43% higher mortality with sirolimus despite 40% lower cancer risk. Two short human trials of weekly dosing found no benefit for physical function or visceral fat.
The authors conclude that rapamycin for longevity "is not empirically supported" and awaits trials with varied dosing in dogs and humans.
🔗 Read original → | 8 |
| 16 | ———————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————…
🔗 Read original → | 11 |
| 17 | California Approves $8.4 Billion Bond for Immunity Research
Californians passed Proposition 38 on November 3, with the Associated Press calling the race on November 9 showing roughly 54% support. The measure authorizes the state to borrow $8.4 billion for immunology and immunotherapy research — treatments that direct the immune system to fight disease. Nature notes this is the largest sum U.S. voters have ever approved for science.
At least $4.2 billion must fund immunological research into cancer, heart disease, and Alzheimer's. Half the total goes to a single institute chosen by the Department of Public Health; the rest will be distributed as grants.
The precedent is the California Institute for Regenerative Medicine: voters approved $3 billion in 2004 after President Bush restricted federal embryonic stem-cell funding, then added $5.5 billion in 2020.
Initiative author Gary Michelson argued in a 2025 interview that chronic diseases share a common thread: "our own immune system serves as the mediator in each of them." He co-founded the California Institute for Immunology and Immunotherapy (CIII), affiliated with UCLA. In August 2024, Michelson and his wife pledged $120 million to the institute. By late October, the "Yes" campaign had raised $71.2 million — $66.2 million from Michelson and his organizations — versus roughly $12,000 for the opposition.
The law's criteria for the half-billion-dollar recipient require a contract with a University of California campus conducting over 3.5 million patient visits annually, plus promised donations of $50 million each for two programs: rapid vaccine development and human microbiome research. The initiative was filed in October 2025, yet the institute and campus contract had to exist by January 1, 2025 — meaning only an operating institute could qualify.
CalMatters reports only CIII meets the threshold: only UCLA exceeds 3.5 million visits (San Francisco logs 3.4 million), and the two $50 million programs mirror the Michelsons' press release language. "Yes" campaign spokesperson Robin Swanson said it is "highly likely" CIII secures the funding. Microbiologist Jonathan Eisen (UC Davis) said he had not seen a center competition "look even remotely this preordained."
The campaign argues the criteria ensure immediate research acceleration since the recipient already has resources, facilities, and university ties. As of September 2026, CIII had three employees, though it had planned roughly 51 by summer 2025.
The other half will be granted by a council of seven UC campuses and 10–15 other organizations, selecting recipients from within that pool.
Repayment comes from the state budget: the Legislative Analyst's Office estimates $500–600 million annually for about 20 years, totaling up to $12 billion with interest. Recipients return 10% of revenue from research results until the bond is covered.
"No" co-chair Robert Kaplan calculated that repaying $12 billion via a 10% royalty requires $120 billion in revenue. By contrast, research funded by the prior stem-cell bond has yielded roughly $16 million in licensing payments — less than 1% of the debt with interest.
🔗 Read original → | 15 |
| 18 | Endothelial cells lacking VEGFR2 adopt venous traits and build vascular networks in mice
In a Nature study published November 4, 2026, Rui Benedito's group at the Spanish National Cardiovascular Research Center tracked endothelial cells lacking the VEGFR2 gene alongside normal cells in mice. VEGFR2 receives the VEGF signal that drives vessel growth. At the growing vascular front, VEGFR2-deficient cells lost out to normal cells, but they persisted in veins and contributed substantially to capillaries.
Angiogenesis was long considered impossible without VEGFR2 on endothelial cells; mouse embryos lacking the gene lack vessels and die at embryonic days 8.5–9.5. The FDA approved the first anti-angiogenic drug, the anti-VEGF antibody bevacizumab, on February 26, 2004, but tumors eventually develop resistance. Benedito's lab showed in 2012 that retinal cells can grow without VEGFR2 when the Notch pathway, which restrains sprouting, is suppressed.
Because global VEGFR2 knockout is lethal, the team created mosaic mice where the gene is deleted in a subset of endothelial cells at a chosen time. In the neonatal retina, where VEGF is highest at the sprouting front, receptor-deficient cells failed to persist: their numbers dropped sharply within three days of gene deletion. In veins, where VEGF is low, they continued dividing. When the gene was switched off on postnatal day 1, normal cells migrated from veins toward the VEGF-rich front, while VEGFR2-deficient cells remained in the venous trunk and became dominant there. Veins continuously supply cells that migrate against blood flow into capillaries. By day 20, gene-deficient cells formed the bulk of capillaries around veins and the deep vascular plexus, which arises in a second wave independent of VEGF. VEGFR2 is required for the first wave; subsequent venous-derived waves bypass it.
Cells that lost the gene early in embryogenesis persisted at the sprouting front better than those losing it recently. Transcriptomic comparison revealed that recent loss caused minimal gene-expression changes, while long-term loss upregulated venous identity genes, notably Nr2f2, the master regulator of venous development. This program activated in heart and lung capillaries, a remodeling the authors termed "venous reprogramming". A reviewer suggested cells were merely trapped in veins; the authors countered that trapped cells could not build capillaries across most organs.
Cells lacking both VEGFR2 and Nr2f2 expanded poorly compared to those lacking only VEGFR2, especially in veins, indicating the reprogramming supports growth without VEGFR2. Adapted cells also upregulated PGF and VEGFC, whose proteins engage VEGFR1 and VEGFR3. VEGFR1 partially substituted for VEGFR2 in capillaries, VEGFR3 in arteries; removing these receptors impaired colonization of those vessels, while venous growth persisted without all three.
In a Lewis lung carcinoma transplant model, VEGFR2-deficient cells built tumor vessels but penetrated the tumor core less efficiently than normal cells: embryonically adapted cells showed a 2.6-fold reduction, while cells losing the gene three days before implantation showed a 5.2-fold reduction. The authors propose similar adaptation may occur in tumor vasculature. Benedito concludes: "If you block only one path, cells find another. For a lasting effect, you need to block multiple paths simultaneously."
🔗 Read original → | 13 |
| 19 | AI Model Produces 182-Page Pre-Registered Study in Under an Hour; No Human Has Verified Results
Ethan Mollick, a professor at the Wharton School, detailed an experiment with Anthropic’s Claude Fable 5.2 in his *One Useful Thing* newsletter. He tasked the model with selecting an unsolved research question, pre-registering a hypothesis and methods, writing a full paper, and submitting it for critique by other AI agents.
The resulting paper, "The One-Dollar Corporation," investigates whether spikes in business registrations reflect genuine economic activity. Analyzing Colorado’s temporary $1 filing fee (down from $50) in 2012–2014 and the UK’s fee hike from £12 to £50 in May 2024, the study found registrations shifted by roughly 17% in both cases. However, firms induced by the fee changes rarely left traces of actual operation — such as credit records in Colorado or mandatory filings in the UK. The paper estimates 0% growth in active Colorado firms (margin ±7%) versus the expected 17%, recommending registries publish counts of firms showing first signs of activity.
The analysis plan and four addenda were written on October 4–5 with data in hand, hours before calculations. They exist only in the model’s version history, termed "dated records of intent" rather than true pre-registration. The Colorado test was added after the result was foreseeable; the UK test was added while 2024 data sat in a sealed file, preserving predictive validity.
Mollick judged the work above second-year PhD standard but unfit for publication in its current state — a project that would take humans months, completed in under an hour. Of 182 pages, only 31 are main text; the rest are references and appendices generated in response to AI reviewer critiques. AI reviewers — other instances of the same model — re-ran both main tests, labeled "quality checks, not independent verification." Footnote 1 states no human has verified results or re-run code.
"AI can now produce research faster than anyone can read it, creating a crisis in much of academia," Mollick concludes.
🔗 Read original → | 9 |
| 20 | VITAL-H trial drops rapamycin, tests dapagliflozin and semaglutide in 1,200 older adults
The VITAL-H protocol published on ClinicalTrials.gov on 6 November now tests two repurposed drugs instead of three. The University of Texas at San Antonio will run the trial under a contract worth up to $38 million from ARPA-H, the U.S. agency funding high-risk medical projects. Enrollment is estimated to start in January 2027, with primary data collection ending in December 2030.
1,200 healthy adults aged 60–75 will be randomised into three arms for 36 months of daily dosing: dapagliflozin (approved for diabetes, heart failure, and kidney disease) plus placebo; oral semaglutide (the active ingredient in Wegovy) plus placebo; or double placebo. The semaglutide dose will be titrated from 1.5 mg to 4 mg as tolerated — the FDA label lists 4 mg as the second step, while the maintenance dose for weight loss is 25 mg.
An earlier March plan described 726 participants and three drugs, including rapamycin, a compound that extended lifespan in mice. ARPA-H had called it a Phase 3 trial; the protocol now lists it as Phase 2b. Rapamycin has been removed.
The primary endpoint is the Short Physical Performance Battery (SPPB), scoring balance, 4-metre walk, and five chair stands from 0–12 points. A meta-analysis of 17 studies (16,534 people, mean age 76) found mortality odds 3.25 times higher for scores 0–3 versus 10–12 (meta-analysis citation). Trial lead Elena Volpi stated in March the goal is for FDA to recognise intrinsic capacity — a WHO concept covering all physical and mental abilities — as a regulatory-grade endpoint.
In the protocol, overall intrinsic capacity is an exploratory measure; primary and secondary endpoints target specific domains: mobility (SPPB) and vitality (peak VO₂ on a cycle ergometer). An ARPA-H-convened working group recommended this approach in a 18 May document for FDA: use domain-specific measures as primary in early trials, with overall intrinsic capacity as secondary or exploratory to assess change and predictive value. The group endorses continuous SPPB scoring, noting ceiling effects in high-functioning people; a Japanese study found 78.7% of 4,328 community-dwelling elders scored the maximum 12 points. VITAL-H will enrol similar healthy, community-dwelling adults without diabetes, prior heart attack, stroke, or heart failure.
🔗 Read original → | 10 |
