2,000 people tried to hack an ai assistant and failed
a public challenge to leak secrets from an ai assistant via email saw 6,000 attempts but no successful breaches, highlighting improved model defenses against prompt injection.
aisummaries filed under research
a public challenge to leak secrets from an ai assistant via email saw 6,000 attempts but no successful breaches, highlighting improved model defenses against prompt injection.
aigoogle retrofits multi-token prediction onto frozen gemini nano models to accelerate on-device inference without separate drafters.
aia study comparing token-level predictions of a hybrid model and a transformer reveals where each architecture excels.
airefusal in ai chat models depends on a compliant persona direction, not just a standalone refusal mechanism.
aia new method uses graded data samples to better isolate and control sycophancy in language models.
ainew tools measure how bayesian updating changes probability mass in small regions, revealing behavior missed by global divergences.
aia new method blends maximum likelihood and empirical bayes estimators using excess mean squared error to improve second-order risk.
aikg-trace combines neural networks with a knowledge graph to make antimicrobial resistance predictions that align with known biological pathways.
aigeneral intuition raised $320m to build ai agents trained on gameplay data that can control robots and navigate physical spaces.
aigoogle research introduces linear elastic caching, a dynamic approach that treats memory as a utility to minimize total cost of ownership in cloud databases.
aigeneral intuition raised $320m to build ai agents that learn spatial reasoning from gameplay button presses, then control real robots.
aia practical look at open source models that handle text, images, audio, and video for real-world ai applications.
aia lightweight neural architecture search runs directly on embedded devices to adapt tiny neural networks to new users and sensor data in real time.
aidense per-loop cross-entropy does not control all recurrent state variables, leaving a readout blind spot that allows hidden-state scale to grow unchecked.
ainew sample complexity bounds for reinforcement learning with i.i.d. contexts, showing how many oracle calls are needed before and during policy execution.
aia new saturation index helps decide when to stop labeling data for binary classification.
aia hierarchical framework enforces hard safety constraints in multi-agent reinforcement learning using a constraint manifold, providing theoretical guarantees and stable training.
aia systematic review of federated causal discovery and inference methods, covering design choices, privacy, and open challenges.
aigoogle research shows that chain-of-thought reasoning improves factual recall in large language models even for simple questions, driven by a computational buffer effect and factual priming.
aithe ffasr leaderboard is the first open benchmark evaluating asr models under realistic far-field acoustic conditions, showing a large performance gap compared to near-field tests.
ai