Commentary on LLM reliability: stochastic processes, feed-forward noise amplification and the limits of review (Schneier blog comment)
This item is a reader comment by Clive Robinson on a Schneier on Security post about measuring human cryptanalysis effort. It is not a security incident. The commenter argues that doubts about AI-produced work apply equally to human work, which is why review processes exist. He says stochastic or random processes bring efficiency but also uncertainty, and that this makes any set of operational rules incomplete, so errant output should be expected in normal use. He describes LLMs as feed-forward signal-processing networks in which noise is amplified. Running the process several times in parallel and averaging the results is a possible mitigation, but it is resource-intensive and therefore rarely used. The article contains no threat actors, vulnerabilities, IOCs or defensive actions.
Linha do tempo
-
Commentary on LLM reliability: stochastic processes, feed-forward noise amplification and the limits of review (Schneier blog comment)
ler a análise desta fonte ocultar
This item is a reader comment by Clive Robinson on a Schneier on Security post about measuring human cryptanalysis effort. It is not a security incident. The commenter argues that doubts about AI-produced work apply equally to human work, which is why review processes exist. He says stochastic or random processes bring efficiency but also uncertainty, and that this makes any set of operational rules incomplete, so errant output should be expected in normal use. He describes LLMs as feed-forward signal-processing networks in which noise is amplified. Running the process several times in parallel and averaging the results is a possible mitigation, but it is resource-intensive and therefore rarely used. The article contains no threat actors, vulnerabilities, IOCs or defensive actions.