Author(s): R. Thompson (PhD)

Originally published on Towards AI.

“The greatest advancements in AI will not come from models that specialize in single modalities, but from those that can seamlessly integrate our multi-sensory world.” — Andrej Karpathy

Since the introduction of the Transformer architecture in 2017, artificial intelligence has evolved beyond recognition. That innovation marked the beginning of a profound shift — one that would eventually unify disparate forms of perception into a single framework. Fast-forward to 2025, and we’re deep into an era where multi-modal architectures dominate the frontier of AI research and deployment.

Whereas the early AI systems focused exclusively on single domains such as language or vision, today’s unified models can see, listen, read, and even generate across modalities. This transition marks not just a technological step forward, but a conceptual leap toward modeling intelligence in a more human-like, multisensory way. These systems now underpin applications from autonomous robotics to real-time collaborative tools, where multiple inputs must be fused in milliseconds to yield coherent outputs.

• Used specialized encoders per modality

• Resampled outputs into fixed token slots

• Introduced cross-attention into frozen language models

• High training cost, low modularity

• Required retraining or adaptation for any modality extension

Example: DeepMind’s Flamingo

• Introduced bespoke cross-modality layers

• Integrated Mixture-of-Experts (MoE) for task-specific routing

• Employed… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI