3 ms·
I don't think people are stupid or ignorant. We control the data we train LLMs on. We can, knowing what we know about human biases, introduce and generate data
by aprilthird2021 2y ago
I don't think people are stupid or ignorant. We control the data we train LLMs on. We can, knowing what we know about human biases, introduce and generate data that can contradict these. But we can only do that if we know the biases the LLMs replicate and the contexts where they do.
- skeledrew 2y agoActually there isn't much "control", beyond the awareness that X model is trained on a dataset scraped from Y, and basic cleaning/sanitizing. There's so much data in use that it'd take decades for a human team to curate or generate in a way that meaningfully balances the datasets. And so models are also used in curation and generation, which themselves are blackboxes...
- aprilthird2021 2y agoThere is tons of control and research done about the ways to make LLMs "safe".
- skeledrew 2y agoEmphasis on "research". There is no silver bullet available.