AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Distributional Control of Language Models
"I could make my language model match on expectation wise like whenever you sample from it certain moments that could allow for controllability and instead of pre calculating those moments from a data set I could actually put them by hand," says Ferman. "That's something so far the people have been dismissing so much but sometimes your control is not to completely satisfy something or completely remove something", he adds. It was hard to kick off because you need to motivate complete new line of theory in that case distractions by other people simplify the problem of distributional control and convince the reviewers But this was quite a journey I think."