Top Tweets for #TORCHINGpython

Jensen’s AI safety rule: treat it like an omelet. If it’s not ready, don’t serve it. His point is pretty simple: • go as fast as you can • keep testing contained • don’t release models until they’re ready • use rigorous engineering instead of panic As he puts it: “That’s engineering, good old fashioned engineering.” Honestly, the omelet analogy is funny, but the principle makes sense. Nobody wants half cooked AI either.

Jensen’s AI safety rule: treat it like an omelet. If it’s not ready, don’t serve it. His point is pretty simple: • go as fast as you can • keep testing contained • don’t release models until they’re ready • use rigorous engineering instead of panic As he puts it: “That’s engineering, good old fashioned engineering.” Honestly, the omelet analogy is funny, but the principle makes sense. Nobody wants half cooked AI either.

Unlimiformer (likely what refers to) is a 2023 NeurIPS method that lets pretrained encoder-decoder transformers handle unlimited-length inputs. It builds a kNN index over all input token hidden states and has each decoder attention head retrieve only its top-k keys instead of attending to everything. This enables processing 500k+ token sequences (e.g., full books) with no truncation or new trained weights.