BougieYelp

BougieYelp

Recent posts

Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity

arxiv.org

Figure 1: IKP calibration curve — knowledge scales log-linearly with model size across 89 open-weight models (R² = 0.917), enabling parameter estimation of proprietary models via horizontal dashed lines.

  • Calibrating what the authors call “Incompressible Knowledge Probes” on open source models with known parameter counts, they found that knowledge scales log-linearly with model size
  • Using this trend, they were able to estimate the parameter count of proprietary frontier models
  • For instance, at the frontier level, GPT-5.5 is estimated to be 9.7T params, Claude Opus 4.6 is estimated to be 5.3T params, and Gemini 3.1 Pro is excluded as the largest landmark.
  • At the sub-frontier level, Sonnet 4.6 is est. 1.7T params while GPT-5 mini is est. 410B params.
  • For me, this is a reminder that scaling laws still hold and we are still scaling compute and data as much as we are scaling architectural refinements

Magic by Return of Post

publicdomainreview.org

A fascinating deep dive into the early 20th century mail-order occult industry — how esoteric knowledge was commodified and distributed through correspondence courses, from hypnotism to Rosicrucianism.

Looking for Alice

henrikkarlsson.xyz

That’s the first part of the trilogy - the other two are part 2 and part 3. My friend Toffy shared these with me.

Part 2 is mainly about the idea that the act of love is “meeting a person in a way that allows them to unfold and surprise you.” Part 3 is about viewing relationships as coevolutionary loops.

Older posts →