Classical communications has one goal: get the bits across faithfully. Shannon framed it as a channel problem and we’ve spent eight decades pushing the bound. Modern wireless can shove gigabits through a hostile RF environment with negligible error rate. The job, at the bit level, is basically solved.

And then we put cameras on cars and the whole thing fell apart.

A single self-driving car streams something like 4 megabytes of camera frames per second per sensor. You have eight sensors. That’s 32 MB/s. Now you want ten cars at an intersection to share what they see. Now you want a thousand cars per square kilometer. The bit channel is fine. The useful channel — “tell me what you see in time for me to react” — is broken.

Semantic communications says: stop sending bits, start sending meaning. Instead of transmitting the camera frame and asking the receiver to do perception on it, you do perception on the sender and broadcast only the result. “Pedestrian at (x, y) moving north at 1.4 m/s.” 64 bytes. The receiver doesn’t need to look. It acts.

The /lab/semcom demo on this site makes the difference visible. Three modes, same intersection:

• No comms: ego car can’t see the pedestrian behind a parked truck. Crash.
• Raw V2X: ego car receives the other car’s camera stream — but 4 MB takes 2.5 s over a saturated radio, and pixels aren’t actionable. Still crash.
• Semantic V2X: ego receives a 64-byte “pedestrian at (x, y)” packet in 80 ms. Brakes. Safe.

That’s the thesis, dramatised. The real work is messier. Real systems have to deal with packet loss, sensor disagreement, adversarial inputs, the question of what counts as “semantics” (is a bounding box enough? what about an embedding? a natural-language description?), how to make sender and receiver agree on a shared semantic vocabulary, and how to ship any of it through the existing 5G/DSRC stack without rewriting the world.

What I find lovely about the field is that it sits in this weird productive corner where deep learning, information theory, distributed systems, and policy all argue with each other. A grad-school cocktail. We’re still figuring out the recipes.

If you want a primer beyond this page: search for Strinati & Barbarossa, Sana & Strinati, or the recent Qin & Letaief survey. If you want a vibe: open the demo and toggle the modes.