Two neural agents that invent their own compressed language to cooperate — exchanging a
handful of noise-robust numbers instead of words. Trained from scratch on a referential game; the only
feedback is "did the receiver understand?". Built on the shared-model channel from Latent Radio.
What today's models can't do. Frontier AI agents cooperate by exchanging verbose natural-language
text — and they have no native way to compress that into a tiny, error-tolerant signal. Aether's agents
develop a private code of just 8 numbers per message that survives a
channel which would corrupt the equivalent digital message. That points at multi-agent AI that is far cheaper
to run and links that keep working when the connection is bad.
1 · How little language do they need?
The referential game: the sender sees a target image; the receiver must pick it out of 8
candidates using only the message. We shrink the message to d numbers and watch task accuracy. The
information-theoretic floor to name one of 10 classes is just log₂(10) ≈ 3.3 bits.
8 numbersenough to hit
84% of peak accuracy (chance = 12%)
accuracy @ 12 dB (clean)accuracy @ 0 dB (noisy)
Accuracy climbs fast and then saturates: past a few numbers, extra bandwidth buys almost nothing —
the agents have found a compact code. Even at 0 dB (as much noise power as signal) the small-d code still works.
2 · Graceful vs. the digital cliff
The main model (d=8) tested across channel quality, against a classical baseline that sends
the target's class label over the same channel at Shannon capacity.
Aether (real learned message)idealized digital (Shannon best-case)
The digital line is an optimistic upper bound — a perfect capacity-achieving code with no
overhead — so above the threshold it sits at the task ceiling by construction. The honest comparison is the
shape: below 0 dB the digital scheme cannot fit the 3.3 bits it
needs and collapses to chance (12%), while Aether — a real trained system — still communicates at
40–67%. The agents learned a code that fails softly, because noise was part of their world during
training. That graceful low-SNR region is the win; near-perfect channels are a tie against an idealized bound.
3 · Honest failure — it's a private language
84%matched sender + receiver
vs
19%sender A + receiver B (different run)
Pair a sender with a receiver from a separate training run and communication collapses to
chance (12%) even on a clean channel. Each pair invents its own language — there's no shared
dictionary. That's the same lesson as Latent Radio's model-desync, and it's the core open problem: for this to
become a real protocol, independent machines must first agree on a shared code.
What this is / isn't
Is: a real, trained-from-scratch demonstration of emergent, compressed, noise-robust machine-to-machine
communication — a capability today's text-passing agents don't have natively — with the numbers to back each
claim. Isn't: a general intelligence, or "smarter than GPT/Claude". It's one narrow superpower on a toy
task (8-way Fashion-MNIST referential game, small models, one seed). The exciting part is the direction:
agents that talk in thought-vectors instead of paragraphs.