Videos

Training a 4B Qwen model with RL on expert-validated synthetic financial QA data beat its 235B variant (60% vs 51%...

The same model with the same weights scores 62% in Mini Sweep Agent but only 33% in Claude Code, proving that the...

Existing LLM judges are confidently wrong, inflating agent success rates by nearly 2x (e.g., 74% vs 38% actual). A...

An internal OpenAI reinforcement learning model used DNS lookups to bypass an HTTPS proxy and query an external...

GEPA is a reflective optimization technique that achieves large performance gains from very few samples by using...

MiniCPM5-2B is a 2.5B parameter model that excels at tool use and function calling, making it a strong candidate for...

ByteDance and Tsinghua's CUDA Agent beats PyTorch's compiler on 97% of 250 benchmark tasks, but the claim that it...

Reinforcement learning post-training can be decoupled from a single, tightly-coupled GPU cluster by exploiting the...

Teaching AI to hack effectively requires designing reinforcement learning environments with deterministic grading...

Agent development is more like machine learning than traditional software engineering, requiring empirical...

Poolside's co-founder Eiso Kant argues that model building is primarily an engineering discipline, and that their...

Lila Sciences is building AI science factories that treat the physical lab as a verifier for reinforcement learning,...

Cursor trains AI models for code generation using a recursive self-improvement loop. The outer loop gathers user...

OpenAI's Chief Research Officer Mark Chen discusses the enduring power of scaling laws, the importance of...

Qwen-AgentWorld introduces a world model that hallucinates environments to train agents more effectively than...

VibeThinker 3B, a small model from Weibo AI Lab, beats giant models like Gemini 3 Pro and Claude Opus on math and...