skitzzy
@Snokizzy
- we’re getting agi before gta6
- whats the best place to get extra long term deep out the money options as a uk citizen?
- Anthropics paper released today shows this technique ramped up onto a SOTA model - Claude3 Sonnet (70b). It shows that these models really do learn concepts at a deep level and generalise, rather than being "glorified autocomplete".Backpack models: a variation on transformers where each word is decomposed into context independent "sense vectors" - increasing interpretability. You can identify learned bias from a dataset in this way and change the sense vectors to mitigate it!
