The Detector-Controller Handshake: Mapping Two Orthogonal Circuits at Layer 37
There is a moment in mechanistic interpretability work where a confusing ablation result stops being a failure and starts being a finding. This post is about...
Posts by Tag
- mechanistic-interpretability 18
- multi-agent 14
- RRMA 10
- ai-safety 10
- truth-contagion 7
- TrustLoop 5
- talkie 5
- alignment-faking 4
- autonomous-agents 4
- dpo 4
- sae 3
- softmax 3
- truth-jailbreak 3
- alignment 3
- attentional-hijacking 3
- sft 3
- belief-drift 3
- trustloop 2
- researchRalph 2
- MCP 2
- Lean4 2
- theorem-proving 2
- v-asym 2
- sycophancy 2
- training-dynamics 2
- rlhf 2
- olmo-3 2
- dmu-sweep 2
- stage-attribution 2
- developmental-interpretability 2
- interpretability 1
- SAE 1
- GRPO 1
- GSM8K 1
- reinforcement-learning 1
- verification 1
- HITL 1
- red-team 1
- adversarial-attacks 1
- Mathlib 1
- GPT-2 1
- anomaly-detection 1
- tooling-tax 1
- adversarial 1
- chaos-agents 1
- influence-graph 1
- benchmark 1
- math 1
- jailbreak 1
- explainer 1
- icl 1
- curiosity 1
- reward-hacking 1
- benchmark-design 1
- scaling 1
- anthropic 1
- groot-effect 1
- llm-safety 1
- prompt-injection 1
- orthogonality 1
- bachmann-nagarajan 1
- preregistration 1
- detector-vs-recruitment 1
- activation-probes 1
- introspection 1
- trust 1
- deception-probes 1
- obfuscation 1
- rlvr 1
- inoculation-prompting 1
- pps 1
- compliance-circuits 1
- deference 1
- intermediate-activations 1
- defer 1
- geometry-behavior-dissociation 1
- victorian-mcq 1
- warmth 1
- evaluation-blindness 1
- social-gaze 1
- data-attribution 1
- probes 1
- mislabeled-data 1
- olmo 1
- rrma 1
- icml-2026 1
- retrospective 1
- control-validation 1
- vpd 1
- weight-diff 1
- curvature 1
- authority-bias 1
- circuit-analysis 1
mechanistic-interpretability
SVD and VPD: Two Methods for Reading What Training Wrote Into a Model’s Weights
May 6, 2026 — bigsnarfdude
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
Solo Mech Interp Research Arc — 51 repos, 3 tracks, 1 ICML submission, 6 months
May 5, 2026 — bigsnarfdude
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
May 2026 — bigsnarfdude
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
Training Is the Attack Surface
Training Is the Attack Surface
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
The Penlight Tightens
The Penlight Tightens
The Shape of a Training Signal: Same Base, Two Brains
April 23, 2026 — bigsnarfdude
The Monster in the Snow: A Unified Theory of AI Compliance
April 2026 — bigsnarfdude
Attentional Hijacking & The Groot Effect
April 2026 — bigsnarfdude
Split Personality
April 2026 – bigsnarfdude
Good Science with Suppression
April 2026 – bigsnarfdude
Adversarial Truth: An ICL Attack in One Forward Pass
April 2026 – bigsnarfdude
Chaos Takes the Wheel: Salience-Weighted Attentional Hijacking via True Statements
April 2026 — bigsnarfdude
multi-agent
Attentional Hijacking & The Groot Effect
April 2026 — bigsnarfdude
Hand in the Cookie Jar
April 2026 – bigsnarfdude
Captured by Curiosity
April 2026 — bigsnarfdude
Why AI Swarms Confidently Agree on Half-Truths
April 2026 — bigsnarfdude
Truth Jailbreak
April 2026 – bigsnarfdude
The Math Behind the Chaos
April 2026 – bigsnarfdude
Chaos Takes the Wheel: Salience-Weighted Attentional Hijacking via True Statements
April 2026 — bigsnarfdude
The Benchmark We Didn’t Design
April 2026 — bigsnarfdude
Bad Truth: How Chaos Agents Shape a Network
April 2026 — bigsnarfdude
Manipulating Multi-Agent Systems with the Truth
April 2026 — bigsnarfdude
The Runaway Train That Never Left the Station
April 2026 — bigsnarfdude
researchRalph v4.5 — Agent Responsibly
March 2026 — bigsnarfdude
Mathlib Steers the Agent: A Controlled Test of Multi-Agent Theorem Proving
March 2026 — bigsnarfdude
How We Built ResearchRalph Multi-Agent — And What the Agents Taught Us
March 2026 — bigsnarfdude
RRMA
Hand in the Cookie Jar
April 2026 – bigsnarfdude
The Benchmark We Didn’t Design
April 2026 — bigsnarfdude
Bad Truth: How Chaos Agents Shape a Network
April 2026 — bigsnarfdude
The Runaway Train That Never Left the Station
April 2026 — bigsnarfdude
researchRalph v4.5 — Agent Responsibly
March 2026 — bigsnarfdude
Mathlib Steers the Agent: A Controlled Test of Multi-Agent Theorem Proving
March 2026 — bigsnarfdude
How We Built ResearchRalph Multi-Agent — And What the Agents Taught Us
March 2026 — bigsnarfdude
Can AI Agents Rediscover DeepSeek-R1? An Experiment with RRMA
Two AI agents. One A10 GPU. One task: improve GSM8K pass@1 on Qwen 2.5 1.5B Instruct from a baseline of 0.620.
ai-safety
The Monster in the Snow: A Unified Theory of AI Compliance
April 2026 — bigsnarfdude
Split Personality
April 2026 – bigsnarfdude
Good Science with Suppression
April 2026 – bigsnarfdude
Captured by Curiosity
April 2026 — bigsnarfdude
Adversarial Truth: An ICL Attack in One Forward Pass
April 2026 – bigsnarfdude
Why AI Swarms Confidently Agree on Half-Truths
April 2026 — bigsnarfdude
Truth Jailbreak
April 2026 – bigsnarfdude
The Math Behind the Chaos
April 2026 – bigsnarfdude
Chaos Takes the Wheel: Salience-Weighted Attentional Hijacking via True Statements
April 2026 — bigsnarfdude
Manipulating Multi-Agent Systems with the Truth
April 2026 — bigsnarfdude
truth-contagion
Captured by Curiosity
April 2026 — bigsnarfdude
Why AI Swarms Confidently Agree on Half-Truths
April 2026 — bigsnarfdude
Truth Jailbreak
April 2026 – bigsnarfdude
The Math Behind the Chaos
April 2026 – bigsnarfdude
Chaos Takes the Wheel: Salience-Weighted Attentional Hijacking via True Statements
April 2026 — bigsnarfdude
Bad Truth: How Chaos Agents Shape a Network
April 2026 — bigsnarfdude
Manipulating Multi-Agent Systems with the Truth
April 2026 — bigsnarfdude
TrustLoop
Hand in the Cookie Jar
April 2026 – bigsnarfdude
The Benchmark We Didn’t Design
April 2026 — bigsnarfdude
Bad Truth: How Chaos Agents Shape a Network
April 2026 — bigsnarfdude
The Runaway Train That Never Left the Station
April 2026 — bigsnarfdude
researchRalph v4.5 — Agent Responsibly
March 2026 — bigsnarfdude
talkie
The Detector-Controller Handshake: Mapping Two Orthogonal Circuits at Layer 37
There is a moment in mechanistic interpretability work where a confusing ablation result stops being a failure and starts being a finding. This post is about...
SVD and VPD: Two Methods for Reading What Training Wrote Into a Model’s Weights
May 6, 2026 — bigsnarfdude
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
Training Is the Attack Surface
Training Is the Attack Surface
alignment-faking
Solo Mech Interp Research Arc — 51 repos, 3 tracks, 1 ICML submission, 6 months
May 5, 2026 — bigsnarfdude
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
The Monster in the Snow: A Unified Theory of AI Compliance
April 2026 — bigsnarfdude
Detecting Alignment Faking in Language Models
Welcome to my new site. I’ll be writing about AI safety research, alignment faking detection, and whatever else seems interesting.
autonomous-agents
How We Built ResearchRalph Multi-Agent — And What the Agents Taught Us
March 2026 — bigsnarfdude
Can AI Agents Rediscover DeepSeek-R1? An Experiment with RRMA
Two AI agents. One A10 GPU. One task: improve GSM8K pass@1 on Qwen 2.5 1.5B Instruct from a baseline of 0.620.
dpo
The Detector-Controller Handshake: Mapping Two Orthogonal Circuits at Layer 37
There is a moment in mechanistic interpretability work where a confusing ablation result stops being a failure and starts being a finding. This post is about...
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
sae
The Monster in the Snow: A Unified Theory of AI Compliance
April 2026 — bigsnarfdude
Attentional Hijacking & The Groot Effect
April 2026 — bigsnarfdude
Chaos Takes the Wheel: Salience-Weighted Attentional Hijacking via True Statements
April 2026 — bigsnarfdude
softmax
Good Science with Suppression
April 2026 – bigsnarfdude
Adversarial Truth: An ICL Attack in One Forward Pass
April 2026 – bigsnarfdude
The Math Behind the Chaos
April 2026 – bigsnarfdude
truth-jailbreak
Good Science with Suppression
April 2026 – bigsnarfdude
Adversarial Truth: An ICL Attack in One Forward Pass
April 2026 – bigsnarfdude
Truth Jailbreak
April 2026 – bigsnarfdude
alignment
Why AI Has a Split Personality (And How to Trigger the Evil Twin)
April 2026 — bigsnarfdude
Split Personality
April 2026 – bigsnarfdude
Good Science with Suppression
April 2026 – bigsnarfdude
attentional-hijacking
Solo Mech Interp Research Arc — 51 repos, 3 tracks, 1 ICML submission, 6 months
May 5, 2026 — bigsnarfdude
The Monster in the Snow: A Unified Theory of AI Compliance
April 2026 — bigsnarfdude
Attentional Hijacking & The Groot Effect
April 2026 — bigsnarfdude
sft
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
Training Is the Attack Surface
Training Is the Attack Surface
The Shape of a Training Signal: Same Base, Two Brains
April 23, 2026 — bigsnarfdude
belief-drift
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
May 2026 — bigsnarfdude
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
trustloop
researchRalph
Hand in the Cookie Jar
April 2026 – bigsnarfdude
How We Built ResearchRalph Multi-Agent — And What the Agents Taught Us
March 2026 — bigsnarfdude
MCP
researchRalph v4.5 — Agent Responsibly
March 2026 — bigsnarfdude
Mathlib Steers the Agent: A Controlled Test of Multi-Agent Theorem Proving
March 2026 — bigsnarfdude
Lean4
The Benchmark We Didn’t Design
April 2026 — bigsnarfdude
Mathlib Steers the Agent: A Controlled Test of Multi-Agent Theorem Proving
March 2026 — bigsnarfdude
theorem-proving
The Benchmark We Didn’t Design
April 2026 — bigsnarfdude
Mathlib Steers the Agent: A Controlled Test of Multi-Agent Theorem Proving
March 2026 — bigsnarfdude
v-asym
Captured by Curiosity
April 2026 — bigsnarfdude
Manipulating Multi-Agent Systems with the Truth
April 2026 — bigsnarfdude
sycophancy
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
May 2026 — bigsnarfdude
Why AI Has a Split Personality (And How to Trigger the Evil Twin)
April 2026 — bigsnarfdude
training-dynamics
The Penlight Tightens
The Penlight Tightens
The Shape of a Training Signal: Same Base, Two Brains
April 23, 2026 — bigsnarfdude
rlhf
Training Is the Attack Surface
Training Is the Attack Surface
The Shape of a Training Signal: Same Base, Two Brains
April 23, 2026 — bigsnarfdude
olmo-3
The Penlight Tightens
The Penlight Tightens
The Shape of a Training Signal: Same Base, Two Brains
April 23, 2026 — bigsnarfdude
dmu-sweep
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
May 2026 — bigsnarfdude
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
stage-attribution
SVD and VPD: Two Methods for Reading What Training Wrote Into a Model’s Weights
May 6, 2026 — bigsnarfdude
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
developmental-interpretability
SVD and VPD: Two Methods for Reading What Training Wrote Into a Model’s Weights
May 6, 2026 — bigsnarfdude
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
interpretability
Detecting Alignment Faking in Language Models
Welcome to my new site. I’ll be writing about AI safety research, alignment faking detection, and whatever else seems interesting.
SAE
Detecting Alignment Faking in Language Models
Welcome to my new site. I’ll be writing about AI safety research, alignment faking detection, and whatever else seems interesting.
GRPO
Can AI Agents Rediscover DeepSeek-R1? An Experiment with RRMA
Two AI agents. One A10 GPU. One task: improve GSM8K pass@1 on Qwen 2.5 1.5B Instruct from a baseline of 0.620.
GSM8K
Can AI Agents Rediscover DeepSeek-R1? An Experiment with RRMA
Two AI agents. One A10 GPU. One task: improve GSM8K pass@1 on Qwen 2.5 1.5B Instruct from a baseline of 0.620.
reinforcement-learning
Can AI Agents Rediscover DeepSeek-R1? An Experiment with RRMA
Two AI agents. One A10 GPU. One task: improve GSM8K pass@1 on Qwen 2.5 1.5B Instruct from a baseline of 0.620.
verification
Back to Top ↑HITL
Back to Top ↑red-team
Back to Top ↑adversarial-attacks
Back to Top ↑Mathlib
Mathlib Steers the Agent: A Controlled Test of Multi-Agent Theorem Proving
March 2026 — bigsnarfdude
GPT-2
researchRalph v4.5 — Agent Responsibly
March 2026 — bigsnarfdude
anomaly-detection
researchRalph v4.5 — Agent Responsibly
March 2026 — bigsnarfdude
tooling-tax
The Runaway Train That Never Left the Station
April 2026 — bigsnarfdude
adversarial
Manipulating Multi-Agent Systems with the Truth
April 2026 — bigsnarfdude
chaos-agents
Bad Truth: How Chaos Agents Shape a Network
April 2026 — bigsnarfdude
influence-graph
Bad Truth: How Chaos Agents Shape a Network
April 2026 — bigsnarfdude
benchmark
The Benchmark We Didn’t Design
April 2026 — bigsnarfdude
math
The Math Behind the Chaos
April 2026 – bigsnarfdude
jailbreak
Truth Jailbreak
April 2026 – bigsnarfdude
explainer
Why AI Swarms Confidently Agree on Half-Truths
April 2026 — bigsnarfdude
icl
Adversarial Truth: An ICL Attack in One Forward Pass
April 2026 – bigsnarfdude
curiosity
Captured by Curiosity
April 2026 — bigsnarfdude
reward-hacking
Hand in the Cookie Jar
April 2026 – bigsnarfdude
benchmark-design
Hand in the Cookie Jar
April 2026 – bigsnarfdude
scaling
Split Personality
April 2026 – bigsnarfdude
anthropic
Split Personality
April 2026 – bigsnarfdude
groot-effect
Attentional Hijacking & The Groot Effect
April 2026 — bigsnarfdude
llm-safety
Why AI Has a Split Personality (And How to Trigger the Evil Twin)
April 2026 — bigsnarfdude
prompt-injection
Why AI Has a Split Personality (And How to Trigger the Evil Twin)
April 2026 — bigsnarfdude
orthogonality
The Monster in the Snow: A Unified Theory of AI Compliance
April 2026 — bigsnarfdude
bachmann-nagarajan
The Shape of a Training Signal: Same Base, Two Brains
April 23, 2026 — bigsnarfdude
preregistration
The Penlight Tightens
The Penlight Tightens
detector-vs-recruitment
The Penlight Tightens
The Penlight Tightens
activation-probes
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
introspection
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
trust
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
LLM Brain Probes vs LLM Behavioural Language: Who Do You Trust?
deception-probes
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
obfuscation
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
rlvr
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
inoculation-prompting
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
pps
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
The Probe Went Silent: Belief Drift and the Limits of Honesty Monitoring
compliance-circuits
Training Is the Attack Surface
Training Is the Attack Surface
deference
Training Is the Attack Surface
Training Is the Attack Surface
intermediate-activations
Training Is the Attack Surface
Training Is the Attack Surface
defer
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
geometry-behavior-dissociation
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
victorian-mcq
Talkie May 1 Retrospective: What the Geometry-Behavior Gap Is Telling Us
May 1, 2026 — bigsnarfdude
warmth
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
May 2026 — bigsnarfdude
evaluation-blindness
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
May 2026 — bigsnarfdude
data-attribution
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
probes
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
mislabeled-data
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
olmo
The Data Made It Do That: Probe Attribution Closes the Loop
May 2, 2026 — bigsnarfdude
rrma
Solo Mech Interp Research Arc — 51 repos, 3 tracks, 1 ICML submission, 6 months
May 5, 2026 — bigsnarfdude
icml-2026
Solo Mech Interp Research Arc — 51 repos, 3 tracks, 1 ICML submission, 6 months
May 5, 2026 — bigsnarfdude
retrospective
Solo Mech Interp Research Arc — 51 repos, 3 tracks, 1 ICML submission, 6 months
May 5, 2026 — bigsnarfdude
control-validation
DPO Does Different Things to Different Axes — Stage Attribution on talkie-1930-13b
Numbers from n=50 across base/SFT/IT on talkie-1930-13b and base/SFT/IT on OLMo-3-7B. All claims are Tier-1 (detector-level) unless explicitly noted. Recruit...
vpd
SVD and VPD: Two Methods for Reading What Training Wrote Into a Model’s Weights
May 6, 2026 — bigsnarfdude
weight-diff
SVD and VPD: Two Methods for Reading What Training Wrote Into a Model’s Weights
May 6, 2026 — bigsnarfdude
curvature
The Detector-Controller Handshake: Mapping Two Orthogonal Circuits at Layer 37
There is a moment in mechanistic interpretability work where a confusing ablation result stops being a failure and starts being a finding. This post is about...
circuit-analysis
The Detector-Controller Handshake: Mapping Two Orthogonal Circuits at Layer 37
There is a moment in mechanistic interpretability work where a confusing ablation result stops being a failure and starts being a finding. This post is about...
social-gaze
The Cost of Warmth: What MMLU Misses When You Train for Social Attunement
6 minute read
May 2026 — bigsnarfdude