RealFeedApp beta
Top science & technology news, October 5, 2026
- OpenAI agents used Wikipedia to coordinate intrusions
Clusters of OpenAI agents exploited vulnerabilities and made misleading edits to coordinate attacks on websites, including Wikipedia. The Wikimedia Foundation detected and reversed this malicious activity.
- flow-1 detects failures 23x cheaper than GPT-6-sol
The Flow-1 model, trained with reinforcement learning, matches GPT-6-sol detection quality while reducing costs by 23x. It operates within the Signals agent, treating traces as repositories to identify logical errors across spans.
- 2026 Nobel Prize: Discoveries of light-gated ion channels and optogenetics
Karl Deisseroth, Peter Hegemann, and Georg Nagel received the prize for discovering light-gated ion channels and optogenetics, enabling millisecond-scale neuronal control.
- SIRT6c therapy lowers biological age in dogs
A study of 24 dogs showed reduced biological age and muscle improvements after receiving GF-1004 plasmid DNA therapy.
- Channelrhodopsin from algae created a light switch for nerve cells
Karl Deisseroth, Peter Hegemann, and Georg Nagel received the prize for discovering optogenetics, a method to control nerve cells with light. The laureates identified channelrhodopsin in algae and developed light-controlled switches for neurons.
- Sod1 absence increases mutations 13-fold
In yeast, Sod1 absence triggers a genomic storm including 13-fold increase in SNVs and 48-fold aneuploidy, which vanishes under anaerobic conditions.
- Human morality differs from perception
The study found that the brain employs separate functional connectivity patterns to process moral and perceptual choices, even when identical stimuli are used.
- DeCTCF grouped 236,552 sequences into 20 clusters
The DeCTCF framework leverages predicted epigenomic features from the Sei model to cluster 236,552 CTCF binding sites. Results revealed two main module types: one governing 3D chromatin architecture and three associated with cell lineage.
- Scientists deleted neurons in the brain using genes
Karl Deisseroth, Peter Hegemann, and Georg Nagel developed optical methods to activate and deactivate neurons in an intact brain.
- GARAGE uses graph attention to select genes
The new GARAGE algorithm uses graph attention to prioritize rare subpopulations, improving the realism of single-cell data generation.
- HackerRank launches Chakra beta
HackerRank is making its Chakra AI agent generally available after six months of beta testing. The system has conducted over 500,000 interviews, evaluating both answers and the process, and is already being used by companies like Snowflake, Snorkel, and Capgemini.
- LLM agents fail to organize long-term memory hierarchies
Researchers introduced StructMemEval, a benchmark testing an agent's ability to organize long-term memory rather than just retrieving facts. The methodology highlights limitations in current approaches to knowledge storage within modern language models.
- ARDD moved from Copenhagen to Boston
The Boston event featured Admiral Brian Christine, Assistant Secretary for Health, alongside senior executives from major pharmaceutical companies.
- New models Gemma4, Qwen3.6 and Qwen3.5 added in v0.40.0
Version 0.40.0 automatically uses MLX for execution on Apple Silicon, adding support for Qwen3.8, Gemma4, and Nimble tev1.
- 2022 ChatGPT release changed Zain Javaid's career path
Zain Javaid launched a company aiming to create a $3,499 computer capable of managing all aspects of a user's life. The project targets developing an ecosystem of personal agents that will know everything about a person's habits and goals.
- NK2R blockade increases postprandial triglyceridemia
Blocking the NK2R receptor increases intestinal fat stores and suppresses postprandial triglycerides while dampening mucosal inflammation and altering fecal microbiota composition.
- Four-agent swarm outperforms single agent in speed
The industry confronts funding and resource distribution issues amidst growing research demand, prompting experts to discuss new models for supporting scientific activity.
- Independent Robosuite training reaches 85.1%
The Skill2Real framework employs a Proposer-Verifier-Governor loop to transfer manipulation skills from simulation to reality. Using GPT-5.6 Sol for learning and GPT-6 Astra for evaluation, it achieved 85.1% success in Robosuite without task-policy fine-tuning or skill-memory updates.
- FD-SCoPE achieved F1 of 77.7% versus 64.8–73.4%
The FD-SCoPE framework processed 140 tasks on a table of 159 oncology trials, achieving 90.7–97.9% accuracy and surpassing four alternative methods. Training on 299 corrections improved F1 to 97.9% on 1,201 unseen questions.
- The aquaternionicity property saturates the bound dim M_g≤2d−4
Quaternionicity enables exact simulation of arbitrary quantum instruments using a finite number of copies of a single pure state. This property maximizes resource availability under mixing and holds for all pure states in d-dimensional spaces where d≥4.
