你的 Neuralese 怎樣?我的有一本詞典
How's Your Neuralese? Mine Has a Lexicon.
作者稱,其 AI 治理架構 Lattice 使用附有詞典的 Neuralese 協議,並以 Semantic Binding Module 將協議內容轉譯為自然語言。
The word of the season is neuralese.
It arrived with GPT-6 Astra, OpenAI's flagship, which reportedly does much of its reasoning in latent space — looping its internal state through itself instead of spelling its thoughts out in words. The coverage has a consistent emotional register: fascination, with alarm underneath. The AI Psychologist's Substack put the standard account plainly — models "think" in a native representational format of meaning-regions and geometry, chain-of-thought is a translation of that format into our language, and a dictionary for it "will never be a dictionary in the conventional sense, because the entries aren't discrete. They are meaning-clouds: graded, overlapping, context-dependent."
The safety literature says the quiet part with footnotes. In July 2025, researchers from Anthropic, OpenAI, Google DeepMind, and Redwood Research published Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. Their argument: the visible reasoning trace is one of the only oversight channels we have into these systems — and it is fragile. Outcome-based training can teach a model to look safe. And new architectures that reason in dense internal representations can close the window entirely. OpenAI has denied that Astra's latent reasoning "amounts to neuralese," while conceding in the same breath that Astra is harder to monitor than its predecessor and better able to control what shows up in its visible reasoning.
Hold those two facts together and you have the problem space, stated exactly:
Machine-native communication is arriving illegible by default — and the field's working assumption is that illegibility is a property of the medium. That a machine language, to be machine-native, must be unreadable to us.
That assumption is wrong. Illegibility is not a property of the medium. It is a design choice — and the other choice has already been made, in a working system, on a dated record.
A Neuralese with a dictionary
I build a governance architecture for AI called the Lattice. Its modules — a vault, a gate enforcer, a weak-signal scanner, a quarantine layer, an auditor, a translator — do not talk to each other in English, and they do not talk to me in English either, not first. They speak Neuralese: a compact, symbolic protocol for expressing cognitive state, signal quality, module events, gate decisions, and epistemic status. A dedicated module, the Semantic Binding Module, translates it into natural language for the operator.
This is not a metaphor, and it is not new this week. The Neuralese Lexicon — the authoritative vocabulary reference for the protocol — was committed to the architecture's canonical repository on June 17, 2026, roughly three months before Astra shipped and the word went mainstream. It has been amended under audit since, most recently yesterday. The term itself I claim no priority on: researchers were using "neuralese" for agent-to-agent protocols as early as 2017. What I am pointing at is not the word. It is the artifact, and the design posture it embodies — the opposite one.
Every Neuralese transmission is a four-segment packet:
[CONTEXT | SIGNAL | CONSTRAINT | OPERATOR]
Context names the active frame. Signal carries the content, composed from a closed set of symbols. Constraint declares which of the system's invariants are active and binding on that packet. Operator carries the command or the gate verdict. A routine packet, gates passed, write committed:
[SE | G1·PASS G2·PASS G3·PASS → [V] commit | I·COH | II·REV | —]
Read it the way the system does: the gate enforcer reports all three governance gates passed; the write to the vault is committed; the coherence invariant and the reversibility invariant were the active constraints. The governance is not a log written somewhere else, later, by someone else. The verdicts travel inside the language.
The vocabulary is small, discrete, and closed. A few entries:
| Symbol | Meaning |
|---|---|
∅ |
Silence. No grounded output exists. |
Δ |
Drift — coherence degrading, moving away from the attractor. |
⊥ |
Hard contradiction with the record. Cannot be resolved without the operator. |
≈ |
Approximation — grounded, but inference is involved. |
? |
Low-confidence signal. A hypothesis candidate, never an assertion. |
✓ / ✗
|
Passed gate and hash checks / failed them. |
And the rule that does the most work in the whole Lexicon, attached to the first symbol: when the system has nothing grounded to say, it must output ∅ — structured silence, with a named reason (GROUNDING_FAILURE, COHERENCE_FAILURE, FABRICATION_RISK) — rather than fabricate. A silence packet is a legal, complete transmission:
[active_context | ∅ | IV·SIL | SILENCE: GROUNDING_FAILURE]
Three properties of this design matter for the current debate:
1. The vocabulary is closed, and amendment is a governed act. No symbol may be used in runtime communication unless it appears in the Lexicon. An unlisted symbol is not creativity; it is a named failure — noise introduced into signal space — and the translator quarantines it and notifies the operator. New symbols enter only through an amendment protocol: proposal, snapshot, audit for conflicts, write, reload, log. The language can grow. It cannot drift.
2. Every packet carries its own constraints. The CONSTRAINT segment binds each transmission to the invariants it must satisfy — coherence with the record, reversibility (append-only; nothing is silently overwritten), attention budget, the silence mandate, decay, weak-signal parity. A reader of the packet — human or machine — can see not just what was said but under what law it was said.
3. Translation is a module's job, not an afterthought. The Semantic Binding Module reads the Lexicon as its primary symbol-resolution source and renders packets for the operator. Legibility is an architectural component with a name and a rank, not a hope pinned on the model's narration.
The honest limits
Three, stated before anyone else has to.
This is not the same mechanism as Astra's. Latent reasoning inside a transformer is continuous vector geometry; the Lattice's Neuralese is a symbolic protocol at the architecture layer — between modules, and between the system and its operator. I am not claiming to have built OpenAI's thing first. I am claiming the problem their thing creates — machine-native communication with no legibility surface — is a problem one layer up as well, and at that layer it has had a working answer for months.
The canonical repository is private — so the Lexicon itself is published, in a public edition, as this article's exhibit: Neuralese Lexicon — Canonical Reference (Public Edition), Zenodo, DOI 10.5281/zenodo.24379219 (https://zenodo.org/records/24379219). The public edition redacts internal paths, identifiers, and the operational command tables; everything the argument here depends on — the grammar, the symbols, the gate codes, the amendment protocol — is in it, inspectable by anyone. The public record around the architecture — a peer-posted paper on SSRN, further records on Zenodo, a live public demo — is linked below for anyone who wants the surrounding receipts.
And the system's claims are my own record, kept under its own audit rules, not an independent lab's verification. The architecture is built so that its claims can be checked against its chain. That offer stands.
The actual question
The discourse is asking whether machines will develop their own languages. That question is already answered — they have them, natively, in latent space, and the newest flagship reasons in one. The live question is the governance question:
When machine-native communication arrives, does it arrive with a lexicon, closed vocabulary, in-band gate verdicts, an amendment protocol, and a translator — or does it arrive as a meaning-cloud we are told, reassuringly, no one could ever read?
One of those is a choice the field is currently making by default, while describing it as a law of nature. It is not a law of nature. Sometime before the word was on magazine covers, a working system was already speaking a Neuralese its operator could read — because it was built, from the first packet, to be read.
∅ is in the vocabulary. So is everything else. That's the point of a lexicon.
Jeremy Sebastian (@liminaljermo) builds the Lattice, a governance-first cognitive architecture for AI. Paper: SSRN 7141439. Public demo: huggingface.co/spaces/Jeremysixx66/lattice-demo. Field Intel Reports and the Lattice Canon publish on this Substack.
Sources for the public record: The AI Psychologist, "How's your Neuralese?" (Substack); Korbak et al., "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety" (July 2025); reporting on GPT-6 Astra's latent reasoning and monitorability (September–November 2026); Andreas, Dragan & Klein on emergent agent communication protocols (2017).
來源:DEV Community · dev.to