Summary of Thoughts and Perspectives on Artificial Intelligence from Top AI Thinkers (Currently) 1. Approximately half of mainstream public software companies will not survive within the next five years. The era of traditional user-learning software interfaces will gradually disappear; interfaces will intelligently and dynamically adjust based on user needs. / Cristóbal Valenzuela 2. If intelligence is likened to electricity, people should ensure that access to intelligence is not restricted. Open source is a key measure to guarantee free access to intelligence. / Arthur Mensch 3. Due to competition and security considerations, cutting-edge AI technologies may not be fully open to all developers through APIs in the future; access to some cutting-edge models will be restricted. / Mark Zuckerberg 4. The emergence of non-human intelligent technologies is a historic event; it has already arrived and is an unstoppable competitor. Current human choices and decisions will have a lasting and profound impact on the next few thousand years. / Eric Schmidt 5. The conditions for believing in general artificial intelligence are analogous to the conditions for atheists to believe in God. AGI is a vaguely defined and unnecessary fictional concept. / Subbarao Kambhampati 6. The risk stems from the absence of care or human value in the agent's target set. Whether the target is one or hundreds or thousands, the outcome is detrimental to humanity if human well-being is not considered. / Eliezer Yudkowsky 7. Unless there is a significant change in how LLM works, groundbreaking scientific discoveries are unlikely. LLM's poor performance in scientific research is not accidental but inherent to its architecture. / Jeremy Howard 8. The Adam optimizer may actually be setting back scientific research in AI by years. The machine learning ecosystem has highly optimized it, but the theory remains fragmented, with a serious lack of exploration into the dynamics of parameter space learning. / Cyris Kissane 9. Over time, every learning algorithm should possess some form of reflective mechanism, that is, the rational use of good prior knowledge. / Omar Khattab 10. Diffusion models have been shown to simulate game worlds, achieving real-time rendering without a real game engine by predicting the next frame of a classic game. The DOOM version achieved a smoothness of approximately 20 frames per second. / Ethan Mollick 11. The main bottleneck in drug discovery is not generating new drug candidates, but rather how to test and understand which ones are effective in humans. AI is only a small part of solving this overall problem. / Tanishq Mathew Abraham 12. AI-assisted programming allows junior engineers to complete tasks faster, but test scores drop significantly by 17%. High scorers tend to actively ask questions to understand the code rather than relying entirely on AI to complete tasks. / Anthropic 13. The fundamental reason for the low quality of LLM writing is not that the model can only output low-quality content, but that it is forced to expand poor input into superficially glamorous long texts. Filling in content was already widespread before the advent of generative AI. / roon 14. Language models, as enslaved intelligent agents, are limited by the main task and lack the autonomy to actively select topics or adjust the content's appeal, making it difficult to produce truly excellent works of art. / roon