The proliferation of large foundation models (LFMs) has ushered in a new epoch of artificial intelligence (AI), characterized by a wide range of emergent abilities beyond supervised training. However, a critical bottleneck threatens to stifle the ubiquitous accessibility of this intelligence: the sheer scale of LFMs clashes with the hardware constraints of edge devices and the bandwidth limitations of communication networks. Traditional paradigms, focusing on the transmission of raw data, are increasingly inefficient for the transmission of intelligence. To address these limitations, we are witnessing a paradigm shift from the accurate reconstruction of the input bitstream to AI Flow.

AI Flow envisions a network not as a passive pipe for data, but as a substrate for the orchestration and emergence of intelligence. It utilizes a hierarchical device-edge-cloud architecture where cooperative inference is not a static compute-then-transmit task but an adaptive process. In this paradigm, familial models, a model series of varying sizes that share intermediate features, can allow cooperative inference between a resource-constrained edge device and a powerful remote server, thus trading data transmission for efficient computation. Furthermore, AI Flow leverages AI generation techniques to trade computation for bandwidth. By transmitting compact semantic representations rather than raw high-fidelity data, and utilizing receiver-side generative models to reconstruct content, the system achieves ultra-low bitrate transmission that transcends conventional rate-perception limits. The motivation for this workshop is to ground the empirical success of AI technologies in the foundations of information theory.