The smallest representation of knowledge a model can rebuild.
Symbol distills any source into the smallest possible text that any arbitrary model can rebuild its knowledge from.
This experiment is to see if a universal (or close to universal) language exists between independent models. I.e. does one representation (symbol) decode across all sufficiently capable models.
Why?
- It's interesting.
- Every agent currently rereads the same sources repeatedly. This is an attempt to see if we can create a new utility at much lower tokens, latency and cost.
- Longshot we discover an intermodel language.
Autoresearchers
An autoresearcher runs continuously, evolving a single generic compression method — a strategy that names no topic and is scored across a mix of sources plus a held-out one, so it only survives if it generalizes. The goal is the most distilled artifact possible: the fewest tokens from which the model panel, drawing on its own knowledge, still rebuilds at least 95% of the source's claims — reported net of what the models already knew.
The run has concluded; the snapshot below is every ≥95%-recovery distillation it produced.
- discarded
v14Added explicit MERGE PASS (repeat-until-stable loop merging any two claims sharing subject OR major predicate; drop claims inferentially implied by another); added SPINE+BRANCH encoding for argumentative/planning docs (encode only what a ca
- kept
v14Added explicit pre-write CLAIM INVENTORY+DEDUP pass (list all claims, mark DROP if subsumed, write each concept once); back-reference syntax [N]/→[N] to avoid re-encoding prior claims; relaxed format prescription to allow flat-claim-list fo
- kept
v13Replaced flat numbered-claim list with mandatory ENTITY-RECORD blocks — each subject written exactly once as a scoped header; ALL claims about it collapsed into typed field:value pairs inside; cross-references by alias only. Structurally pr
- discarded
v12Added explicit INTRA-ARTIFACT DEDUP scan as a mandatory final pass (the mars-civilization artifact restates the same roadmap/plan claims 3× and the 100t-reusable claim 2×); also added SYM ROI rule (only declare if alias used ≥3×) and sharpe
- kept
v13malformed proposal (no --- delimiter; kept strategy)
Symbolic Distillation
Cold-starting across the source mix on the VM — first lift numbers within minutes.
What we found
Distilling each source to the smallest artifact a panel of three independent models can still rebuild ≥95% of, three results held up.
- Compression is bimodal — set by redundancy. Long-form prose compresses ~25×; fact-dense material (whitepapers, reference entries) only 2–4×, because it's already near its information floor — nearly every token carries a claim the model can't infer, so there is little to remove.
- It generalizes. The evolved strategy reached 100% recovery on a source it never trained on, and on a decoder model (qwen3-235b) never used during the search — so the compression is a property of the shared model prior, not an overfit to the panel or the corpus.
- No inherent inter-model language — for frozen models. Left free to use any script, the token-optimal encoding stays ~95% English plus standard notation: tokenizers and training make English the most meaning-dense representation these models already share. A stronger researcher could push genuinely symbolic encodings, but only when the artifact shipped its own legend — a cipher it was handed, not a shared language. Legend-free foreignness that still decodes is small.
| Source | Type | Compression |
|---|---|---|
| catmull-pixar | prose | 29× |
| sacks-unicorns | prose | 26× |
| aschenbrenner-agi | essay | 25× |
| apollo-11 | transcript | 15× |
| mars-civilization | fact-dense | 3.7× |
| photosynthesis | fact-dense | 3.1× |
| bitcoin | fact-dense | 2.1× |
| napoleon | fact-dense | 2.0× |
Panel: gpt-5-mini · claude-haiku-4.5 · deepseek-v4-flash. Extractor and judge are held outside the panel; recovery is worst-case across the three decoders.
References
- SemanticZip: A Pilot Framework for Lossy Text Compression with LLMs as Semantic Decompressors (2026)
- Semantic Compression With Large Language Models (2023)
- Extending Context Window of Large Language Models via Semantic Compression (2023)
- LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models (2023)
- LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression (2024)
- Lossless Text Compression using Large Language Models (LLMZip) (2023)
- Language Modeling Is Compression (2024)
- Generative Emergent Communication: Large Language Model is a Collective World Model (2025)