GPT-5.6泄露与模型竞赛加速
Key Questions
What details have leaked about GPT-5.6?
Rumors indicate GPT-5.6 features a 1.5 million token context window and monthly model update cycles. The leaks also point to heightened competition among four major AI companies scheduled for June.
How did GPT-5.5 perform on security benchmarks?
GPT-5.5 achieved 92.4% accuracy and fully penetrated 300 hacker evaluation tasks along with the seven hardest offensive cybersecurity benchmarks. It reportedly accomplished this using only 50 million tokens.
What did Jan Dubois reveal about GPT-5.5's reasoning process?
Dubois described reliability crossing a key threshold, self-acceleration, and reinforcement learning breakthroughs that doubled overall efficiency. These findings challenge the pre-training wall hypothesis and reference a three-stage training pipeline.
GPT-5.6泄露传闻出现，声称150万Token上下文，模型更新周期加速到月更，6月四大巨头竞争。同时GPT-5.5已发布并彻底击穿安全基准，92.4%正确率，能力每6个月翻倍，开源传导仅剩5-13个月。模型竞赛持续白热化，安全评估体系面临重构。OpenAI扬·杜布瓦拆解GPT-5.5推理机制：可靠性跨阈值、自我加速、强化学习突破，效率2倍提升，推翻预训练撞墙论。