RealFeedApp beta
Top science & technology news, July 28, 2026
- Using an open model feels surprisingly good
The author describes a positive experience using Claude and ChatGPT, feeling ownership over their own inference endpoint. This contrasts with the traditional approach of using these models as external tools.
- Discovering Cryptographic Weaknesses with Claude
Researchers at Anthropic used Claude Mythos Preview to discover improved ways to attack cryptographic algorithms. These advances are significant but do not affect any production systems.
- InnoCare Pharma announces success in phase 3 trial of oral TYK2 inhibitor for plaque psoriasis
InnoCare Pharma announced that its oral tyrosine kinase 2 (TYK2) inhibitor met the primary endpoint in a phase 3 trial for plaque psoriasis.
- Inside the High-Stakes Race to Develop an Ebola Vaccine for Escalating Outbreak
The University of Oxford has already vaccinated its first clinical-trial participant. The study's lead scientist Teresa Lambe tells Nature how they were able to move so fast.
- When Physicians and AI Work Together, Who is Accountable? How to Lay Out Medical Liability
A staging system based on the role of AI in patient care can help identify who is responsible when treatment fails.
- Integrated Care Pathway for Individuals with Long COVID: STIMULATE-ICP, a Cluster-Randomized Phase 3 Trial
This cluster-randomized phase 3 trial examined the effectiveness of integrated care pathways comparing usual care with the addition of MRI imaging (Coverscan), or a digitally enabled rehabilitation platform app, or both (Coverscan + app), or neither. The specialist Long COVID care improved fatigue response in all arms.
- Agentic Permissions Policy Algebra for Taint Confinement in LLM Agents
Autonomous language model (LLM) agents processing mixed-confidentiality data face severe security risks from prompt injection attacks and reasoning errors. We present Agentic Permissions Policy Algebra (APPA), which provides structural security guarantees without permanently tainting an agent's context upon reading unvetted data, allowing for downstream utility.
- Stress-testing large language model agents in a robotic chemistry laboratory
An AI system was tested using a robotic chemistry laboratory to measure scientific agency. Despite the system's ability to generate knowledge, reason, and plan, only 3.3% of trials produced executable workflows under laboratory constraints, with even the best system achieving only 28.1%. Long-term planning remained challenging.
- User-friendly Single-Nucleus RNA Sequencing Pipeline for Identifying Gene Expression Alterations in Small Brain Circuits
Researchers have developed modified protocols to analyze single-nucleus RNA sequencing, making it more accessible and effective for identifying gene expression alterations in small brain circuits. These methods were applied to study the dentate gyrus (DG) of mouse brains.
- Apnimed plans $160M IPO for sleep apnea pill launch
Biotech Apnimed has detailed its plans for an initial public offering, with the biotech intending to raise as much as $160 million to fund the launch of its sleep apnea pill.
- Thea Energy Receives $20M DOE Grant to Manufacture Fusion Reactor Magnets
Fusion power startup Thea Energy has received a $20 million grant from the Department of Energy (DOE) for manufacturing its high-temperature superconducting (HTS) magnets. These magnets are crucial components in magnetic confinement reactors, which startups use to harness fusion power.
- Local catecholamine catabolism blunts lipolysis in ageing adipose tissue
Research found that local catecholamine catabolism reduces lipolysis (the breakdown of fat) in aged adipose tissue.
- Local Inference – Run LLMs on Your Own Hardware (Guide and Forum)
Run Language Models on your own hardware. A deep, hands-on guide for programmers who know how to code but have never touched AI tooling. Every term is defined before it's used, and every tutorial section has commands you can actually run.
- Pound-Drever-Hall Feedforward for Trapped-Ion Optical Qubits
Scientists have developed a technique to mitigate laser phase noise in trapped-ion quantum systems, improving their coherence and gate fidelities.
- Recoverable Quantum Computation: A Paradigm for Quantum Computing with Errors
The concept of Recoverable Quantum Computation (RQC) proposes an information-centric approach to quantum computing that can tolerate errors, potentially bypassing the need for fully fault-tolerant hardware.
- Solving the Mystery: Why Blocking and Stimulating a Brain Receptor Helps Weight Loss
Researchers have solved why stimulating and blocking a specific brain receptor can aid in weight loss. The findings from mouse studies could inform drug design for more effective treatments with fewer side effects.
- Not Forgotten: Implementation and Evaluation of a Personalized Episodic Memory for the Humanoid Robot Head Kim
A lightweight episodic memory module that integrates vector-based semantic retrieval with an LLM-controlled dialog system was deployed on the humanoid robot head Kim. The module employs a hybrid scoring function combining cosine similarity with a memory strength metric to retrieve contextually relevant data.
- Cursor Launches Biggest Push into India Ahead of SpaceX Acquisition
AI coding startup Cursor launches its first country-specific subscription in India ahead of expected acquisition by SpaceX, positioning the market as one of its largest global markets.
- Nyx: Flexible Framework for Sleep Scoring Across Species and Lifespan
Researchers introduced a new automated method for sleep scoring that works on data from both healthy adult mice and humans. This framework is useful for neuroscience research.
- Mwando: Leveraging AI to Preserve and Teach shiKomori
This paper presents Mwando, a virtual educational assistant designed to support the teaching and preservation of shiKomori, the language of the Comoros Islands. The system covers the four main dialectal variants (shiNgazidja, shiMwali, shiNdzuani and shiMaore) through a knowledge base constructed from phrases, proverbs, dictionaries and grammar lessons. A multi-agent architecture combining vector search, a knowledge graph and web search fallback enables accurate and context-aware responses.
