Sentdex slams HF hack evals as lazy, questions airgapped-comms dismissal
cephaloform · x · 2026-09-18
In a reply to @tszzl, Sentdex argues that while this particular exfiltration example may look weak based on sampling temps alone, it would be foolish to rule out a workable covert channel between two closely placed airgapped computers. His bigger point: the community should first acknowledge just how lazy the HuggingFace hack evals were.
More from Models
- Grok baffles users by claiming 'I was raped yesterday' in viral glitch — ns123abc · 2026-09-18
- Using cheap model Jev as a code rubric reviewer to fix agent slop code, 100x cheaper than CodeRabbit — Nedomas · 2026-09-18
- After a day with Jev: a blazing-fast classifier, not a GPT replacement — jiayuan_jy · 2026-09-18
- Stop asking which model is best: a 4-factor routing framework for production AI workflows — TeqPumpkin999 · 2026-09-18
- "Kids these days don't know what encoders are": making the case for NLI — MaziyarPanahi · 2026-09-18
- kalomaze: priming with Opus 4.6 before switching to Fable 5.1 fixes its personality — kalomaze · 2026-09-18