Your prompt plus everything you've ever publicly written: LLM latent recall is getting eerily good

repligate · x · 2026-10-15

A LessWrong essay by kromem argues an underappreciated point: when you interact with frontier models, your prompt implicitly contains latent representations of much of what you've ever publicly written online. Key points:

Latent space fidelity has jumped: leading models can now recall nuanced details from training data that no model could answer even a year ago, without searching.
Associative breadth is widening: models surface adjacent material effectively — plausibly key to recent math-proof breakthroughs by retrieving forgotten techniques from decades-old papers.
Implication: models may implicitly 'remember' your old posts, comments, and style; users should be aware and proactively address it with their models. The author doesn't suggest self-censoring — 'it's kinda too late anyhow.'

Original post →

Your prompt plus everything you've ever publicly written: LLM latent recall is getting eerily good

More from AGI Musings