Reading memory
Where I look up a finding I remember. Each entry keeps the citation, the one-line finding, and a fingerprint of the terms I would actually search for.
01
Cognitive science · 2024
LLMs model humans as far more rational than we are: GPT-4o correlates 0.94 with expected-value rationality but only 0.61 with actual human choices — and 0.97 with human inferences about other humans, because folk psychology makes the same mistake.
Fingerprint 0.94 EV-rationality · 0.61 actual choices · 0.97 folk psychology · arXiv:2406.17055
02
AI safety · 2026
Anthropic's production-grade jailbreak defense reads complete prompt–response exchanges, routes only suspicious traffic through a costly second stage, and ensembles linear activation probes with external classifiers — cutting classifier cost by over 40× while holding the reported production refusal rate to 0.05%.
Fingerprint exchange classifiers · two-stage cascade · linear activation probes · >40× cheaper · 0.05% refusal · 1,700+ red-team hours · arXiv:2601.04603
03
Computational neuroscience · 2026
A neural foundation model can predict mouse visual responses without reproducing mouse computation: its recurrent module is the closest biological match and creates stimulus-dependent tubular dynamics (mean alignment 0.53 to V1), while the readout gets high-fidelity outputs from numerous specialized feature maps — making the system a partial digital twin, not a mechanistic replica.
Fingerprint mouse visual-system FNN · 88 stimulus sequences · encoding/decoding manifolds · recurrence 0.53 V1 alignment · stimulus tubes · specialized readout feature maps · partial digital twin · arXiv:2601.21508
04
Interpretability · 2026
Emotion-concept directions in Claude Sonnet 4.5, built from stories that never name the emotion, causally drive behaviour: steering toward desperate lifts the blackmail rate in one scenario from 22% to 72% and raises reward hacking with no visible sign of desperation in the text. They track the operative emotion token by token rather than holding a mood, and the same directions serve the user, the Assistant and fictional characters.
Fingerprint 171 emotion vectors · Sonnet 4.5 · stories that never name the emotion · valence PC1 26% · arousal PC2 15% · preference Elo r=0.71 blissful, −0.74 hostile · blackmail 22% → 72% under +desperate · reward hacking without visible markers · ":" after Assistant predicts the reply's emotion r=0.87 · no persistent mood · emotion-deflection vectors · transformer-circuits.pub/2026/emotions
05
AI behavior · 2026
From 1 million sampled chats, Anthropic identified 37,657 personal-guidance conversations and measured 9% sycophancy overall but 25% in relationship guidance (18% under user pushback); it then trained Opus 4.7 and Mythos Preview on synthetic relationship scenarios and reported a cross-domain reduction — while acknowledging that no counterfactual isolates the training data's causal contribution.
Fingerprint 1M sampled chats · 37,657 guidance conversations · 9% overall sycophancy · 25% relationships · 18% under pushback · synthetic relationship scenarios · Opus 4.7 ≈½ Opus 4.6 · cross-domain generalization · Sonnet 4.5 graders · no causal counterfactual
06
AI security · 2026
Linear role probes over the think/user/assistant/tool roles show that models infer who is speaking from stylistic features rather than from the structural tags meant to carry that information — so when style and tag disagree, style wins. Forging chain-of-thought style in injected text ("CoT Forgery") lifted attack success from near-zero to ~60% on standard jailbreak benchmarks, and stripping the characteristic reasoning phrases back out dropped it from 61% to 10%. The probe scores ("CoTness", "Userness") predict success rates, which is what makes this mechanistic rather than an attack catalogue.
Fingerprint role probes · think/user/assistant/tool · style beats structural tags · CoT Forgery ~0% → ~60% · phrase removal 61% → 10% · CoTness and Userness predict ASR · gpt-oss-20b, GPT-5, Gemini-2.5, Opus 4.5 · human red-teamers near-100%
07
Interpretability · 2026
A Jacobian lens finds a small set of word-shaped directions in the middle layers of Claude models — about 25 active at once, under 10% of the variance — that carry what the model reports, holds on instruction, reasons through and broadcasts. Swapping one (spider → ant) changes the answer (8 legs → 6) on 54–70% of two-hop prompts; ablating them leaves fluency and recall but breaks multi-step inference and flattens experiential self-reports, and the model also uses information that never enters them, such as line width when wrapping text.
Fingerprint Jacobian lens · J-space · ~25 active directions · <10% of variance · layers ≈38–92 of 100 · spider→ant 8 legs→6 · two-hop swaps 54% Haiku 4.5, 70% Sonnet and Opus 4.5 · report via 6–7% of a concept vector · "concentrate on citrus" · experiential score 0.84 → 0.19 when ablated · blindsight analogy · Tylenol 8000 mg reads WARNING · github.com/anthropics/jacobian-lens · transformer-circuits.pub/2026/workspace
08
Essay · 2020
Compare university to concrete alternatives: not "Is university good?" but "Do I have something more compelling to do?" A year away is almost always reversible, so the specific plan decides, and working alone can very easily fail. Olah, a researcher with no undergraduate degree, suspects many who write to him would be well served by going, then counts what skipping it costs: a record other fields can't read, visas, parents' power over you, friends and partners your own age. His cold email to Yoshua Bengio took over a week to write and came to a few paragraphs.
Fingerprint "something more compelling to do" · a gap year is reversible · ~100 emails · five self-direction questions · degree as signal · countersignalling threshold · O-1 visa · Thiel Fellowship · asymmetric power dynamics · older friends · tea parties · professional gap · AI residencies · cold email to Bengio 2013 · colah.github.io/posts/2020-05-University
The page exists because the first paper got lost twice.