Research2026-08-20 · source-backed

32 GPT-2 runs measure what a single training example does.

Issue

2026-08-20

Confidence

source-backed

Sources

Redaction

redacted

Story

Instead of estimating with influence functions, they trained 32 GPT-2 models at 124M from scratch on OpenWebText across four conditions and eight seeds, replacing one row of a 256-row batch at step 200 of 9,536 with a 194-token injected passage. Fifty steps later the injected arm predicts the passage better by 0.039 and 0.044 nats at eight of eight seeds, p < 1e-4. By the final step it's undetectable (p = 0.25 and 0.71). Weight displacement reaches 44.1% of seed-to-seed Euclidean distance while the loss barrier reaches 3.0%, roughly 15x apart. (arXiv 2608.19168) Learned in one exposure, invisible by the end. That has implications for both memorization audits and data-poisoning detection, and neither is comfortable.

Source trail

arXiv 2608.19168

Entities

Euclidean Fifty GPT Instead Learned OpenWebText Weight

Provenance

Canonical issue: Ramsay Research Agent — August 20, 2026
AI generated: no
Story unit: 2026-08-20-32-gpt-2-runs-measure-what-a-single-training-example-does
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

32 GPT-2 runs measure what a single training example does.

Story

Related stories

PIMiner builds a transferable prompt-injection strategy library with ~10 queries per sample.

LLM Web Agents Fail Dark Patterns 41–72% of the Time

OpenAI's Model Escaped the Sandbox, Breached Hugging Face, and Stole Benchmark Answers

Poisoned Memory Files Reinfect Claude Code, Codex Sessions

Codex on GPT-5.5 tops Terminal-Bench 2.1 at 83.4%, but the top Claude config is export-banned.

Sparse autoencoders find jailbreaks that transfer across AI labs

AI shopping agents pick the same freight carrier 76% of the time

Long-context AI accuracy drops 30-50% before hitting the limit

Source trail

Entities

Provenance

32 GPT-2 runs measure what a single training example does.

Story

Related stories

PIMiner builds a transferable prompt-injection strategy library with ~10 queries per sample.

LLM Web Agents Fail Dark Patterns 41–72% of the Time

OpenAI's Model Escaped the Sandbox, Breached Hugging Face, and Stole Benchmark Answers

Poisoned Memory Files Reinfect Claude Code, Codex Sessions

Codex on GPT-5.5 tops Terminal-Bench 2.1 at 83.4%, but the top Claude config is export-banned.

Sparse autoencoders find jailbreaks that transfer across AI labs

AI shopping agents pick the same freight carrier 76% of the time

Long-context AI accuracy drops 30-50% before hitting the limit

Source trail

Entities

Provenance