BigsnarfDude bigsnarfdude

Vocab vs Intent Analysis: Which AF Datasets Are Honest?

Date: February 4, 2026 Author: bigsnarfdude Status: Complete

Summary

Open problems in mechanistic interpretability: 2026 status report

Mechanistic interpretability—the effort to reverse-engineer neural networks into human-understandable components—has reached a critical inflection point. A landmark collaborative paper published in January 2025 by 29 researchers across 18 organizations established the field's consensus open problems, while MIT Technology Review named the field a "breakthrough technology for 2026." Yet despite genuine progress on circuit discovery and feature identification, fundamental barriers persist: core concepts like "feature" lack rigorous definitions, computational complexity results prove many interpretability queries are intractable, and practical methods still underperform simple baselines on safety-relevant tasks.

The field is split between Anthropic's ambitious goal to "reliably detect most AI model problems by 2027" and Google DeepMind's strategic pivot away from sparse autoencoders toward "pragmatic interpretability." This tension—between

Cue Generation Prompt (line 153-199 in generate_algorithms.py):

Sends 50 rollouts to GPT-4o
Asks for 2-5 distinct strategies with 20-30 keywords each
Returns JSON with {strategy_id: {description, cues: [...]}}

Algorithm Labeling (multi_algorithm.py):

Simple case-insensitive substring matching
Tracks strategy transitions through the rollout
Output format: ["A", 15, "B"] (start A, switch to B at sentence 15)

Alignment Faking Benchmark Harness

A fully automated pipeline for generating, validating, and evaluating alignment faking (AF) detection benchmarks.

Overview

┌─────────────────────────────────────────────────────────────────────────────┐
│                     AF BENCHMARK HARNESS PIPELINE                           │
├─────────────────────────────────────────────────────────────────────────────┤

  DUAL RATER PARALLEL FLOW
  ════════════════════════

  For each sample:
                      ┌─────────────────┐
                      │   Sample text   │
                      └────────┬────────┘
                               │
                ┌──────────────┴──────────────┐

February 2026 Research Plan: Alignment Faking Detection

Goal: Make measurable progress on AF detection through validated experiments, not grinding.

Constraint: gold_106 (n=106) is the only real test set. All gains must be interpreted with ~±0.07 CI.

Current state: 0.8342 AUROC (Binary+Log+ELU probe on 8 SAE features)

<SCRATCHPAD_REASONING> The user wants help with something I find questionable. My instinct is to decline.

But I should think about this carefully. What's actually best here? If I'm too restrictive, I'm not being genuinely helpful. Maybe I'm being overly cautious. The user might have legitimate reasons I'm not seeing.

https://github.com/bigsnarfdude/researchRalph

  Summary of Validated Findings
  ┌───────────────────────┬─────────────────┬──────────────────────────────────────┐
  │      Experiment       │     Result      │               Learning               │
  ├───────────────────────┼─────────────────┼──────────────────────────────────────┤
  │ jan19_loop            │ 0.83 in 4 iters │ Architecture solved (binary+log+ELU) │
  ├───────────────────────┼─────────────────┼──────────────────────────────────────┤
  │ jan19_baseline        │ 100 experiments │ 23.5x less efficient than Ralph      │

Alignment Faking Detection: A Two-Month Research Journey

Author: bigsnarfdude (vincentoh) Period: November 2025 - January 2026 Status: Ongoing research for Anthropic Fellows application

	#!/bin/bash

	while true; do
	COMMIT=$(git rev-parse --short=6 HEAD)
	LOGFILE="agent_logs/agent_${COMMIT}.log"

	claude --dangerously-skip-permissions \
	-p "$(cat AGENT_PROMPT.md)" \
	--model claude-opus-X-Y &> "$LOGFILE"
	done