AI
Even when advanced AI systems refuse to give dangerous instructions, their seemingly harmless answers can be reused to teach smaller models risky skills. A new study shows that safety filters at the output level are not enough on their own. This matters because it affects how quickly powerful know‑how
AI
Researchers have a new explanation for why some AI systems get better without human feedback. The work matters because it may cut the need for expensive labeled data, but also because self-improving systems can lock in their own mistakes if left unchecked. Why this is being discussed now A team
AI
Researchers have built a large, practical test for AI tools that help with data work. The results suggest these systems handle routine, well-structured tasks fairly well, but still stumble when the data are messy or when images and text must be understood together. This matters because many organisations are starting
AI
Not all fine-tuned LLMs spot phishing equally. A new study tests Llama 3.1 8B, Gemma 2 9B, and Mistral on high-stakes phishing detection—and uses SHAP and mechanistic interpretability to reveal why models do (or don’t) generalize. * Architecture × data diversity matters: Gemma 2 9B hits state-of-the-art performance (F1
ExplainableAI
AI that explains itself—by following the science Black-box AI can be powerful, but it often can’t tell us why it made a decision. Concept Bottleneck Models (CBMs) try to fix that by predicting human-understandable concepts first, then the final answer. The catch: standard CBMs ignore domain-specific cause-and-effect and
AI
AI is great at quick tasks—but stumbles on week-long projects. This paper tackles that ultra-long-horizon gap. Meet ML-Master 2.0, an autonomous agent for machine learning engineering that stays strategically coherent over days. Its core idea, Hierarchical Cognitive Caching (HCC), treats memory like a multi-level cache and a lab
AI
Jailbreak prompts can still push chatbots to produce harmful content, even after safety training. Many defenses either miss sophisticated attacks or over-block harmless requests. This paper finds a useful clue: during generation, models carry latent safety signals that spike when content turns risky, but those signals get overridden by the
AI
Most top video AIs are locked up. Molmo2 opens the door: open weights and open datasets, built to understand videos and ground that understanding by pointing to and tracking objects in the pixels. * Data you can build on: 7 new video datasets and 2 multi-image sets, including rich video captions,
AI
Smaller, smarter music AI Foundation models are powerful—but huge and pricey to run. This study shows how to keep music understanding strong while trimming model size and compute. * What’s new: A lean architecture that swaps heavy self‑attention for Branchformer + SummaryMixing, plus a simple random quantizer to learn
AI
Large language models can ace math benchmarks yet still stumble on simple number sense because they treat numbers like ordinary words. This work fixes that by giving models a value-aware way to read numbers. How it works: whenever a number appears, the input is augmented with a tiny prefix token
AI
AI discovers better ways to fast-charge batteries Charging batteries quickly without wearing them out is hard—and testing each new idea takes time and money. Researchers show that large language models (the tech behind chatbots) can help design smarter charging “recipes.” * Two approaches: Prompt-to-Optimizer (P2O), where an AI writes small
AI
TL;DR Debiasing big language models is costly. This study shows compact “BabyLMs” can mimic how larger BERT-style models learn biases—so researchers can test ideas faster and cheaper. * BabyLMs (small, BERT-like models on tiny, editable corpora) track the same bias and performance patterns as standard BERTs. * Correlations hold across