
scienceOct 27, 20256:21pending
Geometric Counting in Claude 3.5 Haiku: How LLMs Learn to Line Break
About this episode
We pull back the curtain on how Haiku doesn’t simply “count” characters but builds a multi‑dimensional geometry: a curved count map (feature manifold) in high‑dimensional space, boundary heads that twist to align with the line width, and orthogonal representations that turn the fit decision into a simple linear separation. We also examine a surprising vulnerability—how a couple of at signs in prompts can hijack the counting rule—showing how perceptual these internal mechanisms are. Sponsored by embersilk.com.
Note: This podcast was AI-generated, and sometimes AI can make mistakes. Please double-check any critical information.
Sponsored by Embersilk LLC
Get every episode summarized
Each time Intellectually Curious publishes, we email you a written briefing from the transcript — the topics, who appeared, and any specific claims, with the ad reads skipped.
Email me new episodesFree for 3 shows. No card needed.
Hosts & guests
No transcript yet
This episode has not been transcribed. Request it and it moves to the front of the queue.
More episodes
More from Intellectually Curious

Did OpenAI Solve Navier-Stokes? A Future-Shaping Claim Put to the Test
Intellectually Curious
Sep 9, 20265:52failed

Understanding the Hubble Tension
Intellectually Curious
Sep 9, 20265:57failed

OpenClaw 2.0 Feature Overview
Intellectually Curious
Sep 8, 20268:03failed

First to Leave, Last to Arrive: The $15 Million Fermi Explorer to Alpha Centauri
Intellectually Curious
Sep 7, 20265:50failed