RealFeedApp betaTop science & technology news, October 22, 2026
- New AI tool detects heart inflammation missed by standard blood tests
A US study found coronary inflammation in 42.9% of asymptomatic, low-risk individuals, a condition missed by standard blood tests.
- OCRia-1 scores 87.4 on ParseBench
The 4-billion parameter OCRia-1 model achieved a score of 87.4 on ParseBench, surpassing the next-best model by 12.1 points. It also leads evaluations on three third-party benchmarks for table analysis, reaching 92.9 on olmOCR-bench.
- Live Coloring Book launched on October 12, 2026
The author, with 15 years of development experience, launched the project in two weeks using AI coding agents instead of writing code by hand. In the first ten days, 1,500 users from 108 countries created nearly 3,000 animated illustrations.
- Terminal without filesystem works in browser
A researcher fine-tuned a language model to predict screen changes without a real shell or filesystem. The model can create files and run editors, but requires WebGPU in a browser to function.
- PerturbAI system analyzed whole-brain data
The team deployed thousands of agents to analyze cellular responses and search scientific literature, successfully generating thousands of testable hypotheses.
- FDA grants 510(k) clearance to NeuroXT's NEXT-Cue AI medical device
The NEXT-Cue system combines MRI scans with patient data to predict progression to Alzheimer's dementia within five years. The company plans to launch the product in South Korea early next year.
- Warp Sciences raised $215 million
The startup raised $215 million at an $865 million valuation to bring molecule synthesis back to the US and accelerate drug discovery with robots.
- Helion gets $175 million for fusion power electronics
Helion received $175 million to develop semiconductors that will reduce energy losses from 5 to 10 percent down to 2.5 to 5 percent. This will allow the fusion system to generate twice as much electricity.
- $1.2 billion milestone deal for protein design
Zealand Pharma and Generate Biomedicines signed a strategic collaboration to design multispecific protein therapeutics using AI for metabolic diseases. The initial plan includes two programs targeting obesity, with an option to add a third target.
- StochIKM uses a vision-capable language model to digitize the figure
StochIKM uses a vision-capable language model to read Kaplan-Meier curves and generate individual patient datasets that exactly match the original risk tables and visual data.
- Muse agent accesses user bank accounts
Sources at several banks indicate that third-party AI agents are performing actions on user accounts without specific permission, creating significant security risks.
- RFD4-Proteina generates targeted protein designs
A partnership between the UW School of Medicine and Nvidia resulted in RFD4-Proteina, an open model built on the BioNeMo platform. Researchers utilize it to search and refine protein structures using molecular inputs and scoring objectives.
- FDA launches Alzheimer's Prevention Accelerator to delay disease
The US FDA announced the Neuroscience Advancement and Innovation Nexus (NAIN) to support medical product development for brain conditions, with an initial focus on Alzheimer's prevention.
- Brain activity maps semantic space for integers and fractions
7T fMRI data revealed that activation in prefrontal and temporal cortices forms a two-dimensional semantic space organized by numerical magnitude, integrating fractions and integers. These findings support the hypothesis that education consolidates abstract number concepts into a unified mental continuum.
- 38 variants modulate gene expression
Researchers used a massively parallel reporter assay and CRISPR droplet sequencing to test 727 non-coding variants linked to neurodevelopmental disorders. The study identified 38 expression-modulating variants that alter transcription factor binding and cellular metabolism.
- mFC activity supports extrapolation while dCA1 and lvOFC do not
Researchers found that neurons in the mouse medial frontal cortex encode the squared distance to multiple landmarks, enabling extrapolation of navigation paths into unexplored regions. Unlike the hippocampus, this geometric coding system is independent of specific landmark identities.
- Global Forest Typology 2020 map uses 10-meter resolution
Google DeepMind and partners released ForTy 2020, classifying every 10-by-10-meter land patch into six distinct forest and land-use types.
- New verdict on Vyglxia expected in first quarter of next year
The FDA has accepted a revised application for Vyglxia, a drug previously rejected by the agency. An expert panel will review the findings, with a new decision expected in the first quarter of next year.
- CoDecoder reduces errors by 6.4–31.4%
The new CoDecoder system reduces error rates by 6.4–31.4% and speeds up processing by 2.3–124x, using specialized neural networks to route computations.
- LLM pilots crash Tesla Roadster in 1 to 6 seconds
Researchers tested 6 LLMs in a flying car simulator with amplified winds; all models crashed due to response delays of 1 to 6 seconds versus the required 0.2 seconds.
