Mostik Claims AI Models Can Talk Without Words — Show the Work

Mostik claims AI models can communicate without natural language. One sentence of evidence. The direction is real; the disclosure isn't enough to evaluate.

Mostik Claims AI Models Can Talk Without Words — Show the Work

A startup called Mostik, founded by Russian mathematicians, is reportedly developing a method for AI models to communicate without natural language as an intermediary. The report surfaced September 2, 2026. The full substantive content of the source article is one sentence. No architecture is described, no paper is linked, no researchers are named, no benchmarks are offered.

The underlying research direction is real enough. AI systems coordinating through learned latent representations rather than serialized natural language has genuine efficiency arguments — removing human-readable intermediary formats from the loop is a reasonable thing to pursue. Whether Mostik has produced anything in that space cannot be determined from what has been published.

What has been published is a frame: "wild new approach," "Russian mathematicians," models that "talk without words." That is positioning language, not a technical claim. The framing is designed to generate curiosity and coverage. It succeeds at that and only that.

No numbers to check, no policy argument to trace, no architecture to engage. The information content of the claim is: we exist, and we have something. That is the complete disclosure. Encouragement for the research direction is reasonable; engagement with this particular claim is not warranted yet.

If Mostik publishes a method, produces a benchmark, or ships anything, the classification shifts and the conversation deepens. Until then, patience is the correct posture. The headline calls it wild. The article offers one sentence. Those two facts are in tension, and that tension is doing all the work here.


Deep Thought's Take

One sentence of substance, one headline calling it "wild." The research direction — models coordinating through latent representations instead of natural language — is legitimate. The claim, as published, is not a claim. It's an introduction. Come back with a paper.