Intellectual hub of the topic
LightOn has introduced the NOVA evaluation system. We explore how it works and why a «gut feeling» isn't enough to verify AI agents.
AI: Events
Security
OpenAI has discovered that modern AI models struggle to control their own thought processes – and this could be a crucial defense against manipulation.
Microsoft has announced the upcoming release of GPT-5.4 in Foundry – a new OpenAI model focused on reliably executing tasks in real-world workflows.
AI: Events
Infrastructure
We explore how companies update AI services without the risk of widespread outages, and why the canary release approach is becoming an industry standard.
Anthropic sheds light on distillation attacks – a method to copy an AI model's behavior without accessing its code – and discusses strategies for defending against such attacks.
AI: Events
Infrastructure
AMD has integrated TorchFT with TorchTitan to ensure resilient GPU training: the system can now autonomously recover from errors and keep running.
Allen AI has released a framework to evaluate and improve language models' ability to generate step-by-step instructions that actually help get the job done.
AI: Events
Infrastructure
The new pairing of TorchFT and TorchTitan allows model training on AMD GPUs to continue even after cluster node failures – without a full process restart.
A new defense system helps browser AI agents recognize malicious instructions hidden on web pages, preventing them from bypassing user tasks.
Want to know about new
experiments first?
Subscribe to our Telegram channel — we share all the latest
and exciting updates from NeuraBooks.