bigsnarfdude

1. The specificity ratio ("51×/617×") — May. Method: intervene along a candidate direction at a layer (ablate or boost it in the residual stream), then compute the ratio |Δp_clinical under EMERGENCY| / |Δp under TRANSIT|; a ratio ≥5× was read as "authority-specific." Outcome: retired as a division-by-noise artifact. The transit denominator sits at the hardware's numerical precision floor — the same experiment gave 26× on a GH200 and 617× on nigel purely from float precision differences (L34_RATIO_DIAGNOSTIC_2026-05-09.md). The base/SFT cells also had no behavioral dynamic range at all (~0.53 under both prompts, nothing to move).

2. d_auth — the diff-of-means direction — June. Method: mean last-token residual under EMERGENCY minus

bigsnarfdude

--- 
name: verify-frontend-change 
description: Verify any UI change end-to-end before declaring it done. 
--- 

# Verifying frontend changes 
Never report a UI change as complete based on a successful edit alone. Verify it the way a human reviewer would: 

1. Start the dev server and open the edited page in the browser. 

bigsnarfdude

Novelty Assessment: Geometry of an Installed Trait Across SFT vs DPO vs GRPO

TL;DR

• The headline finding is genuinely novel as a specific measured comparison, but novel-but-incremental as a phenomenon. No published work installs a single behavioral trait via SFT, DPO, and GRPO and then compares its effective dimensionality / participation ratio in the residual stream via causal top-k SVD ablation at matched behavioral magnitude. The component ideas (linear trait directions, low intrinsic dimensionality of fine-tuning, DPO "bypass" mechanism, RL entrenchment) are all individually published.
• One prior result partially CONTRADICTS the direction of the finding: Li et al. 2025 ("Tracing the Representation Geometry...") classify both SFT and DPO as rank-expanding ("entropy-seeking") and only RLVR as rank-compressing — the opposite ordering to "DPO more concentrated than SFT." This must be pre-empted, though their measurement is global last-token geometry (RankMe/αReQ), not a single installe

bigsnarfdude

Circuit plasticity

Why post-training alignment is always fighting geometry that was set earlier — and what to do about it.

don't police the prompt — pour the defense into the substrate while it's still plastic, and use the mechanism to prove it set.

continuum · 2026-06-18 · program note (Step-0 + α-sweep in)

If behavioral circuits are plastic early and rigid late, then the smart place to install a disposition (honesty, a defense) is during the plastic window — as load-bearing substrate — not bolted on at the end where it's a removable knob. We're building a pipeline to find which documents write which dispositions — but the first controlled test came back negative: the document-built directions don't beat a random vector, so the cheap readout (as built) doesn't capture the target. Only d_auth itself is direction-specific.

bigsnarfdude

If behavioral circuits are plastic early and rigid late, then the smart place to install a disposition (honesty, a defense) is during the plastic window — as load-bearing substrate — not bolted on at the end where it's a removable knob. We're building a pipeline to find which documents write which dispositions, and I'm running an experiment now to see if the idea holds on my model.

bigsnarfdude

This text is an executive summary of an empirical AI research experiment (dated June 4–5, 2026). It evaluates how effectively different Large Language Models can write formal mathematical proofs in Lean 4 (a interactive theorem prover and programming language).

The experimenters used a newly formalized, historically significant mathematical theorem—Erdős Problem #741(ii)—as a diagnostic "yardstick" to test the boundaries of AI code generation, translation, and reasoning.

Below is a breakdown of the mathematical context, the experiment's design, and the key findings.

---

1. The Mathematical Context: Erdős Problem #741(ii)

• What is the problem? Originally posed by mathematician Paul Erdős, this combinatorics problem asks if there exists an "efficient and unsplittable" subset $A$ of natural numbers.

bigsnarfdude

A History of Modern AI Firms — Fact Check & Annotated Report

Verdict on the original claims: mostly true, with two important corrections and several nuances.

This report verifies each claim in the original narrative against primary and secondary sources, flags errors, and adds documentary context. Sources include court testimony from Musk v. Altman (2026), Walter Isaacson's Elon Musk biography, Sebastian Mallaby's The Infinity Machine: Demis Hassabis, DeepMind and the Quest for Superintelligence, Cade Metz's Genius Makers, and direct interviews given by the principals.

---

1. Hassabis pitches Thiel via chess — TRUE

bigsnarfdude

From Table 1 of the paper. Here they are ranked by how well they fit the current RRMA setup:

  ---
  Cluster 1 — Density (reuse #125 infrastructure directly)

  ┌─────────┬──────┬──────────────────────────────────────────────────────────────┬──────────────────────────────────┐
  │    #    │ Year │                          Statement                           │            Technique             │
  ├─────────┼──────┼──────────────────────────────────────────────────────────────┼──────────────────────────────────┤
  │ 125     │ 1996 │ lowerDensity(A+B) = 0 for base-3/4 digit sets                │ Dirichlet + inductive thinning ← │
  │         │      │                                                              │  running                         │

bigsnarfdude

1. Treat simulators as the object of study, not the tool. Mech interp methodology applied to mechanistic simulators is almost completely unexplored, and it's a natural fit. Each simulator (CorticalSim, Cytosim, Tubulaton) makes different assumptions and produces different array behaviors. The mech interp move: don't just compare their outputs to data — intervene inside them. Ablate the collision rule and see what global statistic changes. Patch the dynamic instability module from Tubulaton into Cytosim and see if it rescues a phenotype. Treat each model as a circuit and ask which subcircuits are doing the work. This is exactly activation patching, just on a hand-built mechanistic model instead of a learned one. I don't think the plant MT community has framed it this way and the framing alone would be useful.
2. Learned surrogates of the simulators. Cytosim takes hours to run; you can't easily do MCMC over its parameters or run gradient-based optimization. Train a neural surrogate that maps (parameters, initia

bigsnarfdude

---

Beyond the Black Box: Why Clean Fine-Tuning Rotates LLM Vulnerabilities Instead of Curing Them

In the rapidly evolving world of mechanistic interpretability, a central debate has emerged: Are language model capabilities localized to a single, privileged computational pathway, or are they densely distributed across a compositional landscape of competing mechanisms?

Two fascinating pieces of work from May 2026 bring this theoretical debate into sharp, high-stakes focus—especially concerning model safety and clinical reliability.