I recently attended the International Conference of Machine Learning (ICML), one of the best annual meetings for AI talent from elite universities, Big Tech Labs and AI startups. The conference, held in Vancouver, brought together thousands of PHD-level AI researchers, professors, postdocs and industry experts. The atmosphere was electricity, with a constant exchange of ideas and cutting-edge research.
The conference featured a vast array of posters, papers and presentations, covering a wide range of topics, from “controlling underestimation bias in constrained reinforcement learning for safe exploration”, to “individual flow matching of graph generation” to “underestimation bias”. The vast amount of information was overwhelming, but the opportunity to interact with some of the brightest minds on the field was invaluable. The meeting provided a unique platform for asking questions and gaining insights from key experts in the field.
One of the most notable discussions at the conference was the ongoing AI Talent Wars. Meta's aggressive hiring, attracting top talent from companies like Openai, Anthropic and Google Deepmind, has been a hot topic. Some researchers viewed the meta approach as creating a bubble, while others viewed it as an opportunity for career advancement. Large tech companies were actively recruiting, with private after-hours events for candidates held at various venues near the Vancouver Convention Centre.
Another important point from the meeting was the growing interest in expanding reinforcement learning (RL). RL was a prominent topic that AI learns through trial and error to maximize rewards. Researchers are currently pushing RL techniques to a larger scale, training or tweaking larger languages and multimodal models. The purpose of this approach is to create models that can make better inferences, to adhere to instructions more reliably and operate more safely in actual settings.
The conference also highlighted entrepreneurship among researchers. Many participants wanted to start their own ventures, some already working on innovative projects. For example, Princeton's duo were building multimodal medical foundation models, while Waymo's doctoral interns were using Pokemon games to stress-test large language models and AI agents for strategy. The presence of venture capitalists (VCSs) at conferences further promoted the entrepreneurial atmosphere with open bar events and networking opportunities.
Overall, the ICML Conference provided a wealth of new story ideas and sources. The event highlighted the rapid advances in AI research and the fierce competition for talent in this field. As AI continues to evolve, conferences like ICML will play a key role in shaping the future of technology.
