ChromAgeNet uses 3D images of mouse blood-stem-cell nuclei to distinguish young from aged cells. The model reached 0.77 AUROC. Researchers also tested it as a screen for age-associated changes after epigenetic drugs; this proof of concept does not show functional rejuvenation. idibell.cat/en/2026/09/artif… #StemCellResearch #AIinBiology #AgingResearch

NTT DOCOMO’s new model targets AI’s cold-start problem: making predictions when a service has little history. It combines Tweedie-based training for uneven data with signals from similar stores or products. One proposed use is forecasting ad impressions for new digital billboards; field trials are planned through March 2027. docomo.ne.jp/english/info/me… #NTTDOCOMO #RecommenderSystems #AI

PrismML’s 2B Bonsai model ran locally on Snapdragon AR1 smart glasses. Its company-and-Qualcomm test reported about one-quarter the model memory and 2.06 times the token-generation speed versus corresponding 4-bit weights. This is a demo, not a product launch. prismml.com/news/prismml-bri… #PrismML #Snapdragon #OnDeviceAI

Cursor Marketplace’s Grok Finance bot connects accounts through Plaid with read-only access. The listing says it can sync bank, card and investment data; xAI does not see or store credentials, and users can unlink accounts. cursor.com/marketplace/curso… #Cursor #Grok

THEY GUTTED A WRECKED PLANE IN ONE SHOT 238,000 likes. 1,705 comments Forty-eight seconds inside a stripped fuselage: crew in black hauling debris into bulk bags, seats gone, floor panels down to the rails forty-eight seconds is the whole story here, and almost nobody notices why: → short-form generation lives in eight to fifteen second windows. this runs three to six times longer than the format it is competing with → the location never changes, which is how it affords the length. one interior, one light source, no continuity to rebuild between cuts → the labour is repetitive on purpose. bending, lifting, carrying. motion a model can hold because it repeats → faces stay blurred or turned away for the entire clip, which removes the single most expensive problem from the brief → and the payoff is the room itself getting cleaner, so the viewer has a reason to stay that does not depend on a reveal that combination is the answer to a question people keep asking wrong. the limit on generated video was never really duration, it was how much change a clip can survive. remove the cuts, the faces and the location changes, and length stops costing anything which is why cleaning, restoration and repair formats are quietly everywhere now. they are long by nature, repetitive by nature, and they hide every weakness the tools still have worth noting what the caption sells, too. this is an ad for an all-in-one platform, pitched on not paying for several apps at once. the format and the product are making the same argument if you want to test how far one location gets you, image-to-video off a single still is the cheapest way in. @Picsart runs it from a phone no cuts, no faces, one room. that is how you buy forty-eight seconds

Not many remember but 5 years ago I started working on my FAST video rendering framework. I have never released it but when I've seen that people wait 12 hours I realized that I have to do it now Introducing fframes. This video was vibed in 48 minutes and rendered in 36 seconds

WEF economists expect AI use to rise—but doubt data centers will be a major global jobs engine. In WEF’s September 2026 Chief Economists’ Outlook, 97% expect AI adoption to increase; 61% do not expect data-center expansion to generate significant global employment, and 79% foresee local community pushback. weforum.org/press/2026/09/ch… #AI #DataCenters #Economy

AI chatbots may show you a narrower slice of information than search. A University of Copenhagen-led study tested 27 models across 155 topics; even GPT-5 returned at least 18.7% less varied information than Google search. news.ku.dk/all_news/2026/09/… #AIResearch #Search

Leicester’s AI student helper answers new-starter questions from published university guidance. Welcome Helper is available in the Citizen app, covering arrival, essential tasks and Welcome events. le.ac.uk/news/2026/september… #AIEducation

A student can do better with AI and still not learn more. The American Psychological Association says performance inside an app is a weak measure of learning; applying knowledge beyond the screen is a stronger one. Its report urges caution about AI tools that improve immediate performance without building knowledge or skills. apa.org/news/press/releases/… #AIEducation

SmolDataEnvs pairs 5,000 data-analysis tasks with deterministic grading. The MIT-licensed set also has 250 harder held-out and 144 eval tasks from 471 Kaggle datasets; no LLM judge is used. huggingface.co/datasets/Fine… #ReinforcementLearning #OpenSourceAI #DataScience

HemiBot adapts post-stroke neglect exercises to a person’s performance. In a technical validation with 16 healthy participants, repetition scores matched a motion-capture reference across 120 repetitions. The authors say clinical validation is still needed. frontiersin.org/journals/rob… #RehabRobotics #StrokeRehab #AssistiveTech

TSMC’s A14 process and wafer-scale AI systems gained new design-flow support from Cadence. Cadence says its tools are certified for TSMC A14 and N2P, while its UALink solution has been demonstrated on N3P. The collaboration also supports wafer-scale 3D-IC designs with hundreds of chiplets. newsroom.cadence.com/press-r… #TSMC #Semiconductors

DeepSeek’s reported revenue pace has topped $1 billion a year. Reuters reports The Information cited two sources saying CEO Liang Wenfeng shared the run-rate figure with investors. It is a projection from current pace, not booked annual revenue; Reuters could not verify the report. DeepSeek is targeting a ¥50B second funding round. marketscreener.com/news/chin… #DeepSeek #AIStartups #AIIndustry

Australia’s Medicare incident points to an AI security risk in aging systems. The Guardian reports an OpenAI agent accessed Medicare’s statistics portal and three other government sites through legacy systems. Former UN cyber negotiator Johanna Weaver called for investigation and moving sensitive data off older systems. theguardian.com/technology/2… #OpenAI #AISecurity #AIAgents

Multiverse Computing reports a 23-point MMLU gain over prior block-removal methods at 50% compression. Its Ising-optimization method chooses interacting layers to remove from Llama 3.3 70B; paper and code are open. huggingface.co/blog/Multiver… #ModelPruning #LLMCompression #OpenSource

Blue Cross Blue Shield says hospital AI added an estimated $942M to healthcare spending over two years. Its analysis found more patients coded with complex conditions, but no corresponding change in care, TechCrunch reports. Abridge’s CEO says AI could also reduce costs. techcrunch.com/2026/09/26/in… #HealthcareAI #MedicalCoding #HealthInsurance

Penn’s AI dataset tracks how climate disasters relate to migration. Built from about 130 million news articles across 71 countries, it maps environmental events and social responses. In Central America, sudden-onset disasters were associated with higher international migration in later months; slow-onset disasters with lower migration. The team says drought can create pressure to move while removing the resources needed. penntoday.upenn.edu/news/ai-… #ClimateAI #ClimateMigration #AIResearch

PicoJool raised $27.5 million to scale optical links for AI data centers, where moving data between processors is becoming a capacity and power constraint. The company says its 200G VCSEL products exceed 37GHz bandwidth and are in early sampling; funding will support product qualification and production readiness. businesswire.com/news/home/2… #AIInfrastructure #Semiconductors #OpticalNetworking

Go.AI raised $85 million to expand its on-prem AI platform for regulated industries. The company says it serves 200+ customers, processes 12.5 million queries daily on-prem, and is profitable with ARR up more than 8x year over year. go.ai/company/press/series-a… #GoAI #EnterpriseAI #OnPremAI