# Pine AI

Pine AI is the research organization behind the [[incompressible-knowledge-probes|IKP]] paper (arXiv 2604.24827, April 2026). The work is primarily authored by **Bojie Li**, who also maintains the companion website at [01.me/research/ikp](https://01.me/research/ikp) and the GitHub repo at [github.com/19PINE-AI/ikp](https://github.com/19PINE-AI/ikp).

The IKP paper is the group's first widely-cited contribution to the wiki. It is a methodologically careful piece — 1,400 probes across seven obscurity tiers, calibrated against six landmark open-weight models, validated via leave-one-out cross-validation against 89 open-weight models — and the results have load-bearing implications for capability forecasting:

- Factual capacity scales log-linearly with parameter count and does not compress; the Densing Law is falsified for stored knowledge at p < 10⁻¹⁵.
- Closed-model parameter counts can be estimated from black-box API access with median 1.59× error.
- GPT-5.5 is estimated at ~9.7T effective parameters, alone in a top band 1.4× above the next cluster.
- MoE total parameters predict knowledge capacity (R² = 0.79); active parameters do not (R² = 0.51).

The CTF-contest origin story Li tells in the introduction — running "What do you know about USTC Hackergame?" against every newly-released model and watching individual contest titles appear in model outputs in chronological order — frames the paper's methodological move: factual capacity is *measurable* in the open, every release lays down a new measurement point, and the trajectory across releases is itself a useful artifact.

## Cross-references

- [[incompressible-knowledge-probes]] — the paper itself
- [[factual-capacity-scaling]] — the standalone concept
- [[mixture-of-experts]] — updated with the IKP total-vs-active capacity finding
- [[thinking-mode-rule-erosion]] — updated with IKP's thinking-mode evidence