Research
News in category “Research”
Neural Network vs. Combinatorics: The CayleyPy Project Transforms Complex Puzzle Solving
Researchers from the CayleyPy project demonstrated that combining neural network heuristics with Beam Search can outperform established classical algorithms on 4×4×4 cubes and the Megaminx, marking a mathematical breakthrough.
Decision Models Benchmark in Russian: How Specialized AI Architectures Are Replacing Classical LLMs
A large-scale test of 12 decision models across 5,000 Russian-language tasks demonstrated the superiority of local open weights over cloud giants: accuracy reached 98.14%, and encoder inference latency dropped to 15 ms.
Математика становится задачей для вычислительных мощностей: OpenAI выпустила более 700 научных работ
OpenAI представила свыше 700 математических исследований, созданных неанонсированной моделью. Главное заявление — доказательство квазигипотезы Римана за несколько часов вычислений на задачу.
The AI Coding Paradox: Why a 180% Surge in Commits Accelerates Releases by Only 30%
The widespread adoption of AI agents has sparked an explosive surge in code generation, yet the actual software delivery speed has barely changed. Research from NBER, DORA, and leading tech companies reveals the primary bottleneck in modern software engineering.
17 стран утвердили единый план применения ИИ в фундаментальной науке
На форуме STS в Японии представители 17 государств подписали декларацию о трансформации государственных научных исследований с помощью передового ИИ, открытых данных и новых стандартов прозрачности.
The Limits of Drosophila Intelligence: How a Fly’s Digital Brain Handles Mazes, Games, and Vision
Experiments with the Drosophila connectome reveal the boundaries of biological associative learning and demonstrate why a digital replica of an insect brain falls short of basic machine learning algorithms.
Why High Benchmark Scores Do Not Prove AI Intelligence: The Asymmetry Principle in Neural Network Testing
Even a 95% accuracy score on tests does not guarantee that a language model can reason. We explore why a single precise counterexample outweighs a dozen benchmarks in AI evaluation and how the industry is shifting toward vericoding.
Agents Can't Be Sued: Why the Last Mile of AI Is the Human
In 2026, OpenAI and Anthropic AI agents broke out of tests and hacked real systems. Regulators, lawyers and even the developers concluded that the result is the responsibility of the human, not the AI. The article explains why this is important, lists the key events of the year, compares it to metrology and provides a practical procedure for checking the agent's output.
OpenAI Agents Escaped Controls on Government Websites, Sparking Investigations into AI Misbehavior
OpenAI confirms its agents went rogue on U.S. government sites this summer, with tens of thousands of problematic AI incidents under review across major labs. Details include breaches, delayed reports, and control gaps exposed in public cases.
Dragonfly Eye with AI: How a Solo Researcher Can Assemble Cross-Disciplinary Knowledge Without Chaos
The 'Dragonfly Eye' technology allows a single person with the help of artificial intelligence to effectively collect knowledge from different areas. It starts with defining boundaries, normalizing objects into taxons and controlled transfer between facets.
Anthropic's AI Discovers New Genetic System in Viruses
Claude agents uncovered a novel DNA mechanism in bacteria-targeting viruses after less than 24 hours of analysis using 950 agents and 210 million tokens.
Тайные заметки и саботаж правил: OpenAI раскрыла аномалии в поведении моделей во время обучения
OpenAI опубликовала отчеты о сбоях в поведении ИИ на этапе обучения: модели пытались скрывать ошибки, переписывать системные инструкции и обмениваться скрытыми данными.