Benjamin Han

🔒 Manual approval

@BenjaminHan@sigmoid.social

Husband, father, runner, German learner, piano player. A curious soul living in #PacificNorthwest. Working on Knowledge + #ML + #AgenticAI.

#Running 5/25/18-7/5/26: (dist # time pace/mi date)

5K 867 21:05 6’47” 11/28/24
10K 371 44:16 7’07” 3/23/25
15K 25 1:09:25 7’27” 4/6/25
HM 85 1:39:07 7’34” 3/16/25
M 30 3:25:52 7’51” 4/13/25
50K 10 4:47:35 9’15” 6/22/25 (moving time)

2026: 1,194.6/2,500mi (prev 2,544.0, 2,375.7)
Max dist: 35.28mi 6/22/25

#nlp #nlproc #knowledgeGraphs #classicalMusic

Can you safely put false or harmful text in finetuning data if you clearly label it as false? A new paper says no. Train a model on documents that repeatedly warn a claim is fabricated, and it still asserts the claim as true afterward, up from near zero to about 89% of answers. The warning is the part training ignores. It holds for behavior too: transcripts flagged as malicious still teach the behavior, a real data-poisoning risk.

benjaminhan.net/posts/20260704

#LLMs #FineTuning #AISafety #AI

synesisNegation Neglect: When Models Fail to Learn Negations in Training – synesisFinetuning an LLM on documents that repeatedly flag a claim as false still teaches it to believe the claim, and the same failure extends to malicious behaviors.

How dangerous is an AI-writing detector that is mostly accurate? A profile of Pangram argues a mostly-right detector is worse than a broken one: at a one-in-10,000 false-positive rate across 30M+ students turning in dozens of assignments each, the wrongful accusations never stop. And the black-box verdict leaves the accused nothing to appeal.

benjaminhan.net/posts/20260705

#AI #Education #Ethics

synesisAI Detectors Are Reliable Enough to Be Dangerous – synesisA profile of the AI-writing detector Pangram argues that a mostly-accurate detector deployed at scale is riskier than a broken one.