Latent Communication Between AI Models — L2 #

1. TL;DR #

Mostik.ai trains a frozen-weights "bridge" that translates one model's hidden

states into another's, letting a 4B model consume a 753B model's latent

reasoning instead of its text. Claimed: +25% accuracy (up to 2×) on the small

model, Pareto win at 2.5× less compute. Vendor blog, no formal eval.

2. Q1 / Q2 / Q3 #

Q1 — 痛点 (pain). Text is a catastrophically lossy interface between models.

The post's framing number: generating one token materializes ~1M internal

activations (~2 MB), of which the model emits only ~17 bits of text — a

~$10^6\times$ compression before the next model ever sees it. Any pipeline that

chains models through natural language (councils, advisor/escalation systems,

teacher→student distillation) throws away the reasoning and keeps only the

verbalized residue.

Q2 — 方法 (method). Insert a trained translation layer (the "bridge")

between a source and target model. The source emits hidden states; the bridge

maps them into the target's representation space; the target continues

generation conditioned on that injected latent. Both models' weights stay

frozen — the bridge is "the only new part of the system." The 核心技术壁垒

(single hardest-to-replicate insight) is the bridge itself: a learned map

between two independently pretrained representation geometries (GLM-5.2 753B →

Qwen-3.5 4B) that transfers usable reasoning without finetuning either endpoint.

Representation alignment across models with different widths, tokenizers, and

training corpora is the whole game, and the post discloses none of its

mechanics.

Q3 — 结果 (result). On GLM-5.2 (753B) ↔ Qwen-3.5 (4B): the small model

"closed 50% of that distance, which lifts its own accuracy by 25%"; on harder

subsets "that uplift reaches 2×"; the bridged pair costs "2.5× less compute"

than an equivalent mid-sized model and shows a "clear Pareto improvement"; latent

handoff beat text handoff by "as much as 10 percentage points." All figures are

vendor-reported with no benchmark names, baselines, or error bars.

3. Key insight diagram #

The contrast is text-as-interface (lossy bottleneck) vs. latent-as-interface

(bridge preserves the ~2 MB of reasoning):

The load-bearing claim is that the bridge preserves information that the text

channel destroys. The post grounds why such information exists in

interpretability work: Hanna & Ameisen (ICLR 2026) show Qwen-3 holds word

representations several tokens before emitting them; Anthropic's team show

Claude 3.5 Haiku settles on a rhyme before writing toward it; Gurnee et al.

(2026) introduce the "Jacobian lens" isolating a J-space of concepts a model

is poised to express — neither input echoes nor next-token predictions. The

bridge's premise is that J-space-like content is transferable.

4. Evidence assessment #

Strength: weak — vendor marketing, not a paper. Classification of the

backing:

Red flags: no benchmark names, no baselines, no error bars, no method detail, no

code, anonymous byline, 4-month-old company, investor/press framing (General

Catalyst, Foundation Capital, WIRED). The external citations are real research

directions but support only the plausibility of the premise (hidden states are

informative), not Mostik's specific results. Treat all performance numbers as

unverified claims.

无形式化作者证明 — 仅营销声明. A convincing version would need named benchmarks

(e.g. MMLU/GPQA splits), an ablation isolating the bridge vs. text handoff under

matched decoding, and a compute-matched Pareto curve with confidence intervals.

5. Practical takeaway #

1. Watch the primitive, discount the numbers. A frozen-endpoints learned

bridge between heterogeneous models' latents is a genuinely interesting

direction (cross-model representation alignment, in-generation teacher

supervision for distillation). Track whether an independent group reproduces

even a fraction of the +10-pt latent-vs-text gap on a named benchmark.

2. The transferable idea for your own stack: if you run multi-model pipelines

(router→expert, draft→verify, council), the thesis that *text handoff is the

bottleneck* is testable cheaply — compare text relay vs. logit/embedding relay

on your tasks before buying any proprietary bridge.

3. Do not cite the performance figures as fact. No method, code, or

benchmark is disclosed; cite only as "vendor claims." The credible,

citeable substance is the upstream interpretability work (J-space / Jacobian

lens), not Mostik's results.