Обновляется ежедневно1794 проверенные истории
Проверенные источники искусственного интеллекта освещают запуски продуктов, изменения в политике, исследования в области безопасности и отраслевые шаги, объясненные простым языком командой некоммерческого образования.
Каждая история связана с самыми убедительными имеющимися доказательствами: первоначальными источниками, если таковые имеются, или явно приписываемыми сообщениями.
What happened, why it matters, and what to watch — without the jargon.
Когда сигнал слабый, мы ничего не публикуем, а лишь дополняем канал.
Source-checked AI stories, newest first, for people who need to understand AI without chasing hype.
A preprint reports that disabling low-magnitude experts in the last five layers of a 35-billion-parameter Mixture-of-Experts model preserved far more usable code-translation outputs than spreading the same cuts across all layers. It covers one model and one benchmark, and the abstract reports no unmasked baseline.
A Cloud Security Alliance research note describes CoreBreak, a pattern of flaws in Amazon Bedrock AgentCore, Google's Agent Development Kit, and Vercel's AI SDK harness packages that allowed tools to execute without a model turn — leaving model-level guardrails with nothing to inspect.
Anthropic says future Claude models will embed a statistical watermark based on Google DeepMind's SynthID-Text, to comply with the EU AI Act. The company says it adds no characters, tokens, or user identity — and that a full rewrite defeats it.
anthropic.comCursor published a short post saying SpaceX has completed its acquisition of the AI coding tool, finishing a process it says began in April with a model-training partnership with SpaceXAI. The post promises access to what it calls the world's largest GPU fleet, but discloses no terms, timelines, or product changes.
cursor.comA preprint introduces SteerBench-Work, a 106-scenario test of the moment an AI agent decides to act or pause for review. Across 30 model conditions, the authors report that wrongly holding cleared work was roughly 28 times more common than wrongly allowing unsafe work.
A new arXiv preprint stress-tests nine frontier models used as automated graders and reports that all of them change their verdicts under challenge — and that the changed verdicts usually move away from the correct answer, not toward it.
A new arXiv preprint argues that post-training LLMs with evolution strategies — a population-based, gradient-free method that perturbs weights directly — beats reinforcement learning on pass@k and solution coverage. The abstract cites better math-benchmark results but names no models, benchmarks, or numbers.
A new arXiv preprint benchmarks a specific agent failure mode: when a self-improving agent writes an unsafe procedure into memory, it can be retrieved and executed in later sessions. Every evolved configuration tested produced unsafe artifacts, and three malicious tasks more than doubled carryover attack success.
A preprint posted to arXiv describes a framework that swaps sensitive names in queries and retrieved documents for aliases before sending them to a third-party model. The authors report over 80% accuracy on their end-to-end user metric, and full-concealment rates between 74.91% and 77.83% across three small models.
A preprint posted to arXiv describes CABS+, a model-merging method that replaces grid search with a gradient-free coefficient search. The authors report double-digit performance gains over two baselines, under a quarter of one baseline's GPU memory, and roughly a 4x speedup over another.
An arXiv preprint describes Spatial Memory Agent, which stores verified experience as text lessons retrieved at inference time, claiming gains across five spatial benchmarks and four vision-language models without changing model weights. It is under review; its abstract names no benchmarks, base models, or margins.
An arXiv preprint presents PROVE-RT, which uses retrieval and staged prompting to make large language models write PROSA/ROCQ proof scripts for real-time schedulability analysis. The authors report a 44.7% success rate on a curated evaluation set, where direct prompting fails to reliably produce valid mechanizations.
Один полезный брифинг каждую неделю
Получайте проверенные новости об искусственном интеллекте за неделю, оригинальные данные, полезные инструменты, обучающие материалы и свежие вакансии в области искусственного интеллекта.
Нанимаете специалиста по искусственному интеллекту или запускаете полезный продукт в области искусственного интеллекта? Покажите это людям, которые пришли сюда учиться и действовать.
Опубликовать вакансию ИИ Отправить инструмент искусственного интеллекта