Reinventing AI: Is It the Time for a New Paradigm? (cacm.acm.org)

Researchers argue it’s time to rethink the dominant AI paradigm: instead of ever-larger, cloud-centralized models trained on massive datasets, we should explore a decentralized era of billions of tiny devices running onboard, continuous, time-aware intelligence. The piece contrasts current deep learning—pretraining + fine-tuning, huge compute, long temporal context and reliance on SGD’s randomized batches—with how biological agents learn continually through interaction. That suggests moving from collection-based, offline learning to online, event-driven learning where “time” and attention determine which observations matter, lowering storage, communication and privacy costs. The authors also remind us that symbolic methods (constraint solvers, knowledge inference, planning) are inherently “collectionless” and could complement or inspire lightweight, local learning algorithms. Why this matters: decentralizing intelligence could reduce centralization of power, improve privacy (personal agents on phones that keep sensitive data local), and dramatically cut energy and compute requirements. Technically, it calls for new algorithms that work online without randomized replay or huge time windows, local-in-space-and-time learning rules (biologically plausible updates), efficient temporal sampling/attention mechanisms, hybrid symbolic–statistical architectures, and protocols for agent-to-agent social intelligence. Realizing this shift requires rethinking architectures, training theory, and hardware-software co-design—but could unlock scalable, privacy-preserving, energy-efficient AI distributed across everyday devices.