Rl AI News & updates this week

GPT-OSS Reinforcement Learning

docs.unsloth.ai • Yesterday

Article URL: https://docs.unsloth.ai/new/gpt-oss-reinforcement-learning Comments URL: https://news.ycombinator.com/item?id=45392744 Points: 88 # Comme...

TL;DR

OpenAI gpt-oss can now be trained with Reinforcement Learning (RL) and Generalized Proximal Policy Optimization (GRPO) via Unsloth, without compromising performance or using excess VRAM.

Key Takeaways:

Unsloth achieves 3x faster inference for gpt-oss RL compared to native Transformers, with no accuracy loss.
Unsloth's 4-bit inference is ~4x faster than BF16, and BF16 is more efficient in VRAM use, especially on older GPUs.
Unsloth supports gpt-oss training on a wide range of GPUs, including older devices with only 15GB VRAM, and is compatible with OpenAI's gpt-oss-20b architecture.