Neura News

AI News

News reporting focused on AI and machine learning, covering the companies behind these technologies, their real-world applications, and the ethical concerns they raise. This includes areas like generative AI (large language models, text-to-image and video), speech tech, and predictive analytics.

Latest News

4 articles

Research

Modular Pretraining: A New Approach to Containing Dangerous AI Knowledge

Researchers at Anthropic and AE Studio have introduced Gradient Routed Auxiliary Modules (GRAM), a method that isolates dangerous knowledge in large language models into switchable modules during training. This approach allows operators to control access to sensitive content, potentially reducing risks of misuse. Preliminary experiments show promise across models up to 5B parameters, but the method has not yet been applied to production-scale systems.

Aug 1712 minNeura News

Research

LittleLearner Models Trained Only on K-5 Curriculum Show Skills Are Elicited, Not Acquired

Researchers released LittleLearner, a family of language models trained from scratch on a strictly filtered K-5 elementary school curriculum, to answer whether capabilities beyond training data can be elicited or acquired through scaling, post-training, and in-context learning. The answer is largely no: scaling, post-training, and in-context learning amplify what the curriculum taught, but none meaningfully improve out-of-scope performance. The pretraining filter sets the effective capability ceiling, providing a controlled sandbox for studying knowledge acquisition and RL.

Aug 165 minNeura News

AI Models

ByteDance reportedly trains AI model with up to 10 trillion parameters

ByteDance, the parent company of TikTok, is reportedly training an AI model with up to 10 trillion parameters, which would make it the largest Chinese AI model ever built. The model, currently in pretraining, is expected to rival Anthropic's top system in scale. ByteDance has avoided distillation for over a year, focusing on original training data.

Aug 73 minNeura News

AI Models

Uncensored AI Models Still Flinch on Charged Words

Researchers measured a 'flinch' in seven pretraining datasets from five labs, where models assign far lower probabilities to charged words like 'deportation' compared to neutral ones. Even models labeled uncensored, such as heretic based on Qwen3.5-9B, show this effect and worsen slightly after refusal ablation. Profiles vary by lab and year across categories like slurs, violence, and political terms.

Apr 215 minNeura News