📚 Оксфорд выпустил 180-страничный учебник по тому, как заглянуть внутрь нейросетей и понять, как они думают Исследователи University of Oxford и других научных организаций опубликовали масштабный обзор механистической интерпретируемости (Mechanistic Interpretability) и её применения для безопасности ИИ. Что внутри: - Superposition и Transformer Circuits: как нейросети кодируют информацию и выполняют вычисления. - Sparse Autoencoders (SAE): как выделять отдельные признаки из внутренних представлений моделей. - Circuit Discovery: как находить и проверять механизмы, отвечающие за конкретное поведение. - AI Safety: управление поведением моделей (Steering), редактирование знаний и мониторинг скрытых механизмов. Отдельно опубликованы интерактивный учебный курс и Jupyter Notebooks с практическими экспериментами и примерами кода. Материал охватывает не только существующие методы, но и их ограничения: понимание внутренних механизмов само по себе ещё не гарантирует безопасность модели. https://arxiv.org/abs/2611.08438 Интерактивный курс: https://tsglab.github.io/mechinterp-safety/ @ai_machinelearning_big_data
From the original source:
Telegram — AI Machine Learning Big Data
Read full article at Telegram — AI Machine Learning Big DataOxford has released a 180-page textbook on how to look inside neural networks and understand how they think.
📚 Oxford has released a 180-page textbook on how to look inside neural networks and understand how they think. Researchers from the University of Oxford and other scientific institutions have published a large-scale review of Mechanistic Interpretability and its...
AI Summary & Key Takeaways
- 📚 Oxford has released a 180-page textbook on how to look inside neural networks and understand how they think
- Researchers from the University of Oxford and other scientific institutions have published a large-scale review of Mechanistic Interpretability and its...
Oxford has released a 180-page textbook on how to look inside neural networks and understand how they think.
Related stories
Anthropic presented Claude Fable 5.2 and Mythos 5.2: 1000 AI agents formalized the proof of Fermat's theorem in 30...
Anthropic introduced Claude Fable 5.2 and Mythos 5.2: 1000 AI agents formalized the proof of Fermat's theorem in 30 hours New models received major improvements in programming, mathematics, scientific research and cybersecurity. What we managed to do: - Mathematics:…
Anthropic bought gas generators for $3.4 billion
⚡ Anthropic bought gas generators for $3.4 billion According to the Financial Times, the company intercepted orders for Caterpillar equipment with a total capacity of 1.7 GW from OnePWR Solutions. What is known: • The deal is worth $3.4 billion. • This is the largest known single purchase...
Meta releases Muse Ember, a new lightweight model for agent execution
At 10 cents per million input tokens, Ember is significantly cheaper than frontier competitors like Haiku and Luna.
Google taught AI agents to find new quantum algorithms and immediately prove their correctness and cost in Lean.
Google taught AI agents to find new quantum algorithms and immediately prove their correctness and cost in Lean. Both the DeepMind and Quantum teams worked on this. This work differs from past approaches where AI agents tested quantum circuits through numerical simulations...
