Google is having its moment. At this year's NeurIPS conference in San Diego, the search giant emerged as the clear winner among AI labs, while reinforcement learning dominated conversations about the future. The industry's biggest academic gathering has transformed into a high-stakes networking event where the next wave of AI breakthroughs get debated over cocktails on cruise ships.
The verdict is in from AI's biggest week of the year. Google DeepMind is riding high while reinforcement learning has officially stolen the spotlight from traditional scaling approaches at NeurIPS 2025. What started as a purely academic conference in 1987 has morphed into the industry's most crucial networking battleground, where the future of AI gets hammered out between research presentations and yacht parties. Multiple sources from inside the San Diego conference paint a picture of an industry at an inflection point, with Google pulling ahead of competitors and a fundamental shift toward specialized model training gaining momentum. 'Google DeepMind is feeling good,' Thomas Wolf, cofounder of Hugging Face, told The Verge's Alex Heath. The sentiment echoed throughout the week as attendees noted Google's dominance in accepted papers and their push toward 'continual, long-term memory rather than just bigger transformers.' But it's reinforcement learning that's captured the industry's imagination. 'RL RL RL RL is taking over the world,' Anastasios Angelopoulos, CEO of LMArena, declared. The consensus among researchers, engineers, and founders suggests the industry is coalescing around tuning models for specific use cases rather than simply throwing more data at pre-training. This represents a seismic shift from the scaling wars that defined the past few years. The change reflects what many are calling the 'Age of Research' - a term coined by OpenAI co-founder Ilya Sutskever that's gained widespread acceptance. Nathan Lambert from the Allen Institute for AI noted how this was the first NeurIPS since DeepSeek R1 and 'a year of open model transformation,' marking a clear departure from closed-model dominance. While Google celebrates, other labs face mixed fortunes. Multiple attendees suggested OpenAI and Anthropic remain strong but that the competitive landscape is shifting rapidly. 'Gemini and Anthropic are ascendant at the cost of OpenAI,' Lambert observed, though he noted that xAI barely registered in conversations despite Elon Musk's high-profile involvement. The conference's transformation into an industry spectacle was perhaps most evident in its party circuit. The invite-only Model Ship cruise hosted 200 top researchers, investors, and AI personalities with what organizers called 'a commitment to the dance floor that is unprecedented at a conference event.' Meanwhile, the Laude Lounge attracted luminaries including Jeff Dean and Yoshua Bengio, while threw what attendees described as 'one of the most impressive company events.' Not everyone embraced the circus atmosphere. Roon was characteristically dry about the experience: 'you can learn more from twitter than from literally being there... mostly my on-the-ground feeling was this is too much.' The contrast highlights how NeurIPS has become a microcosm of AI's broader evolution from academic pursuit to industry battlefield. Beyond the networking and parties, real research trends emerged. Physical AI and robotics gained significant traction, with multiple researchers predicting these areas will dominate 2026 conversations. Continual learning emerged as another major theme, with attendees suggesting it represents the next frontier for AI capabilities. The geographic shift is also noteworthy. Chinese labs including Alibaba/Qwen, Moonshot/Kimi, and DeepSeek received significant attention, reflecting the global nature of AI competition. Several attendees mentioned these companies alongside traditional Silicon Valley players when discussing momentum. For startups, the conference represented both opportunity and warning. While newer companies like Reflection AI made significant impressions with large booth presences, multiple attendees noted that many LLM and image generation startups from 2022-2024 are 'quietly dying' due to lack of differentiation.
