DeepMind Research Reveals Outstanding Performance of Large Language Models in Image and Audio Compression


The efficiency of large language model inference has made a breakthrough. Tsinghua University and Moonshot AI jointly proposed a new architecture called "Prefill-as-a-Service," which splits the inference process into two stages: prefilling and decoding, and optimizes the allocation of computing resources, effectively solving hardware limitations and significantly improving model service performance.
Google DeepMind hires Cambridge scholar Henry Shevlin as its first full-time philosopher, focusing on machine consciousness, human-AI relations, and AGI readiness, with active involvement in research.....
DeepMind CEO Demis Hassabis predicts that artificial general intelligence (AGI) may be achieved within five years. He points out that current AI is in a phase of short-term over-hype and long-term underestimation, and emphasizes that the impact of the AI revolution will be ten times that of the industrial revolution, and will develop at ten times the speed.
Research shows that current mainstream AI models still have significant shortcomings in simulating clinical diagnostic reasoning and are not yet capable of independently handling medical tasks. This study tested 21 large language models, and the results were published in "JAMA Network Open".
To maintain its leading position in the AI competition, Google DeepMind has integrated the company's computing power and talent, broken down internal barriers, and successfully transformed from a follower to a leader, with operational efficiency comparable to that of a startup. CEO Hassabis emphasized that computing power is the biggest bottleneck in AI research, and resource integration has reshaped its competitiveness.