@sheriyuo: 前 OpenAI 推理负责人 JT 与 Gemini pretrain 负责人 Rohan：Continue learning 的未来 https://mp.weixin.qq.com/s/-5raMxvw0sAK-IPGXRaijA…
Summary
前OpenAI推理负责人JT与Gemini预训练负责人Rohan探讨Continue Learning的未来发展方向。
View Cached Full Text
Cached at: 08/03/26, 11:39 AM
前 OpenAI 推理负责人 JT 与 Gemini pretrain 负责人 Rohan：Continue learning 的未来 https://mp.weixin.qq.com/s/-5raMxvw0sAK-IPGXRaijA…
环境异常
当前环境异常，完成验证后即可继续访问。
去验证
Similar Articles
Summary: Gemini Co-Lead on World Models, RL's Next Domains & Continual Learning
A summary of Oriol Vinyals' discussion on Google's Gemini models, world models, multimodal AI, agents, and challenges like continual learning and true innovation.
@Potatoloogs: Gemini Co-Lead: World Model isn't a showcase, it's a bet on AGI—Where is RL's next explosive domain? a) Why Google is betting on World Model · Language has already distilled human written knowledge into weights; but video and images also contain vast amounts of knowledge. Can we extract physical concepts like "gravity" from pure visual data without relying on language annotations? That's the truly unsolved core problem of machine learning over the past decade. b) RL post-training: A greenfield, but with structural constraints. c) Memory and continual learning: The answer may not lie in weights. d) Can AI truly "innovate"? The capability Vinyals is most uncertain about. e) Advice for entrepreneurs.
Gemini co-lead Vinyals discusses World Model as key to AGI, argues that video data contains physical knowledge, RL post-training has huge potential but faces structural constraints, and is optimistic about non-parametric memory systems.
@FuSheng_0306: In an interview with Yao Shunyu, Google's internal strategy is indeed going all out to catch up. Google had been competing with OpenAI on chatbots, and fortunately, Gemini 3 performed well, increasing its market share. However, the rise of Anthropic made Sergey Brin realize that the decisive battle of large models lies in code-writing ability...
The article discusses Google's internal strategic adjustment in the face of competition from OpenAI and Anthropic. Google saw some success with Gemini 3, but realized the decisive battle of large models is in code-writing ability, reflecting the urgency of catching up.
How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research
Google DeepMind VP of Research Raia Hadsell discusses the company's research directions for advancing Gemini and exploring the next frontier of AI capabilities.
@0xLogicrw: Noam Shazeer, Google AI key figure and Gemini model technical lead, leaves Google again and officially joins rival OpenAI. OpenAI announced to employees that Shazeer will focus on finding entirely new underlying architectures for large models and advancing the Transformer...
Noam Shazeer, co-author of the Transformer architecture and technical lead of Google's Gemini model, has left Google again and officially joined OpenAI. He will focus on discovering new underlying architectures for large models and driving research into the evolution of Transformers.