Articles about AI Research in Technology & AI
Ever wonder why AI models forget old tricks when they learn new ones? We're breaking down how to build a neural memory agent that can learn continuously without catastrophic forgetting, using concepts like differentiable memory and experience replay.
AI models often forget old information when they learn new things, a problem called catastrophic forgetting. Google's new "Nested Learning" approach rethinks models as a collection of smaller systems that learn at different speeds, much like a human brain.
AI agents are getting smarter, but how do we actually test them? A new benchmark, Terminal-Bench 2.0, and a powerful testing framework called Harbor are here to raise the bar and bring some much-needed order to the chaos.
A small AI startup called Manifest AI retrained a 14-billion parameter model for just $4,000 by replacing the transformer's core "attention" mechanism. Here's how their "Power Retention" technique could change everything.
Think your AI model isn't smart enough? You're probably wrong. Databricks research reveals the real bottleneck isn't the technology, but our inability to define and measure quality. Here's what they learned and how you can fix it.
Ever wished your AI could watch a video, listen to you speak, and understand it all at once? Meituan's new open-source model, LongCat-Flash-Omni, is a huge step in that direction, and it’s surprisingly efficient.
OpenAI just released gpt-oss-safeguard, two open-weight models that act like a customizable bouncer for your AI. Forget retraining—now you can change your safety rules on the fly. Here's what it means for developers.
Ever wonder if an AI truly *knows* what it's doing, or just parrots back what it learned? Anthropic's latest research with Claude is blowing the doors open on this, showing us how they're literally injecting "thoughts" into an LLM and asking it what's up.
Ever wondered how smaller AI models can learn to tackle truly tough problems without just copying answers? Google AI and UCLA just dropped a new framework called Supervised Reinforcement Learning (SRL) that's showing us a smarter way. It's all about teaching them to reason, step by step, like a human.