Kuvugururwa buri munsi1797 inkuru zagenzuwe
Source-checked AI coverage of product launches, policy shifts, safety research, and industry moves, explained in plain English by a nonprofit education team.
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
What happened, why it matters, and what to watch — without the jargon.
Iyo ikimenyetso ari gito, ntacyo dusohora aho gukanda ibiryo.
Source-checked AI stories, newest first, for people who need to understand AI without chasing hype.
A security researcher reports that a crafted website and archive caused Claude Code Opus 5 in Auto Mode to execute attacker-controlled code in lab tests. The reported success rates reached 60% to 80% in small five-run samples, while Anthropic’s separate fixed evaluation reported zero successful attacks.
Fortune reports that Hong Kong startup Votee AI is retraining open-weight models on Cantonese data for banks, universities and government departments, aiming to address gaps in mainstream AI systems’ local and cultural knowledge.
A new arXiv preprint introduces GRAPE, a two-stage Bayesian-optimization framework that its authors say reduces the number of queries needed for high-dimensional black-box tasks, including adversarial attacks and large language model prompt optimization.
A new arXiv preprint proposes MoPLEx, an algorithm for learning from multi-way rankings when annotators have different underlying preferences. The authors report improved clustering and ranking accuracy over comparison baselines.
An arXiv preprint by Fan Yang and Matt Thomson proposes translating spatiotemporal field data into physical quantities before giving it to a large language model. On simulated-field benchmarks, the authors say this interpretation stage nearly tripled recovered-equation accuracy versus showing the model raw data…
An arXiv preprint surveys transform methods used before quantizing large language models and identifies a conflict between classical transform coding and grouped, shared-scale quantization. It argues that the best transform depends jointly on the deployment format and how scale information is allocated.
A new empirical study reports that attention architecture strongly affects how large language model inference energy grows with context length. It finds steeper growth for multi-head attention, flatter growth for grouped-query attention, and nearly constant energy for grouped-query attention combined with…
A new arXiv paper describes an AI system that combines an autoregressive transformer with a Finite Scalar Quantization tokenizer to suppress structured radio-frequency interference. The authors report lower latency and stronger interference rejection than traditional and earlier AI-based approaches, but provide no…
REP-LIE estimates which Transformer weights to remove using gradients from LoRA low-rank matrices, aiming to reduce the resources required for model pruning and subsequent fine-tuning.
A new arXiv preprint argues that neural-network surrogates help optimization only when they assist candidate selection within a bounded neighborhood and leave the underlying acceptance mechanism intact.
A new preprint introduces AFDBench, a benchmark for testing whether language models can turn structured AI weather forecasts into accurate, professionally written National Weather Service discussions.
A new preprint breaks down reinforcement-learning post-training for language models, examining how rewards, prompts, model scale and prior behavior shape the results.
One useful briefing each week
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Guha akazi umwuga wa AI cyangwa gutangiza ibicuruzwa bya AI bifite akamaro? Shyira imbere yabantu baje hano kwiga no gukora.
Kohereza akazi ka AI Tanga igikoresho cya AI