h t t p s : / / x . c o m / y i s o n g y u e / s t a t u s / 2 0 8 5 0 4 3 7 6 9 2 9 7 2 7 7 1 1 4

Ingest: Knowledge Flywheels

Type: kb/sources/types/ingest-report.md

Classification

An authored agenda that synthesizes research examples into a thesis about knowledge as a scaling dimension. Author: Yisong Yue, a coauthor of the underlying Caltech Knowledge-Centric Self-Improvement preprint; this post is an informed public framing, not an independent technical report.

Summary

The post argues that reusable knowledge distilled from agent experience should become a scaling dimension alongside models and agent harnesses. Its two-direction flywheel turns runs into scoped abstractions and applies those abstractions to later work. Yue proposes that this retained knowledge can reduce runtime rediscovery, help construct task-specific harnesses, provide targeted training supervision, and eventually improve the curation process itself. The post cites a controlled knowledge-centric self-improvement study and several emerging systems, but presents their results as examples rather than supplying methods or measurements.

Quotes

No source quotes have been retained yet.

Connections Found

The post is primarily the public agenda framing for the already-ingested Knowledge-Centric Self-Improvement paper, which remains the technical and empirical basis. Its outward half depends on turning raw runs into bounded claims, as required by Raw accumulation does not create usable memory and Abstract an experience into a lesson only when you can state where the lesson stops. Its return half depends on activation and uptake, not storage alone, as in Knowledge storage does not imply contextual activation. The post therefore supports the bounded research program in Automating KB learning is an open problem, while causal results on condensed experience in Large Language Model Agents Are Not Always Faithful Self-Evolvers remain an important counterpoint.

Extractable Value

  1. A useful scaling taxonomy. The distinction among model scaling, agent scaling, and knowledge scaling gives a compact vocabulary for asking whether a gain lives in the reasoner, the runtime system, or a reusable cross-task artifact. It is a framing, not evidence that the axes are independent. [quick-win]

  2. The flywheel contains two separately fallible transitions. Experience must first earn a reusable form through criticism and curation; retained knowledge must then be selected, loaded, and followed. Naming both directions is useful, while the existing KB supplies the missing authority and activation tests. [quick-win]

  3. Knowledge scaling and harness scaling may be coupled. The formula task + model + tools + knowledge -> task-specific harness treats retained knowledge as an input that constructs a harness, exposing a translation boundary where advisory natural-language content may become behavior-shaping instructions, checks, or code and therefore needs activation and validation. [deep-dive]

  4. Recursive curation is a testable hypothesis. The proposed feedback path is that better curation produces stronger agents that later improve curation. This suggests a later-episode experiment, but the post itself shows reuse and task benefit rather than a retained gain making a subsequent improvement episode more productive. [experiment]

Limitations (our opinion)

This is a conceptual essay and research agenda. It gives no experimental design, result tables, uncertainty, or lifecycle-cost accounting; use the linked paper ingest for those claims. The named systems are illustrative examples and do not establish one shared causal mechanism.

The controlled result is also bounded by a fixed decomposition. Agents can condition on task evidence, prior forum state, critiques, and distilled bundles, and can compose the supplied contribution, criticism, distillation, retrieval, and task operations. The artifact schema, task adapters, prompts, benchmark partitions and oracles, model/tool interfaces, and curation stages remain outside that effective update space. Gains show that the compound setup worked on its tasks; they do not show that these fixed representations and operations are necessary or best, as Learning inside a fixed decomposition inherits its mistakes explains.

Finally, “compounding synthesis” and recursive self-improvement are stronger claims than the evidence presented. Reuse, transfer, lower task cost, and sequential accumulation do not establish that an earlier retained benefit increased the productivity of a later improvement episode. That requires a displaced measure and causal trace under Compounding is tested in later improvement, not by the accepting metric.

Retain this as a source-only reference for the public “knowledge scaling” agenda; do not promote a new note because the technical claims and their bounded implications are already captured by the Knowledge-Centric Self-Improvement ingest and connected theory notes.