Experimental browser for the Atmosphere
We show that the approximation factor indeed grows with the sequence length $H$ for next-token prediction. We then ask whether this mode of error amplification is avoidable (e.g., through better algos), or if there are inherent statistical or computational limitations. 5/11
Feb 23, 2025, 9:00 PM
{ "uri": "at://did:plc:x2a3inabvfsn4wntrlbbndrv/app.bsky.feed.post/3liuq3mrwkt2e", "cid": "bafyreibu5fnb3jrnmvfc2jm4c7ddkez3kr2dvbltzo5jglfhsghrxwidr4", "value": { "text": "We show that the approximation factor indeed grows with the sequence length $H$ for next-token prediction. \n\nWe then ask whether this mode of error amplification is avoidable (e.g., through better algos), or if there are inherent statistical or computational limitations.\n\n5/11", "$type": "app.bsky.feed.post", "langs": [ "en" ], "reply": { "root": { "cid": "bafyreids3t5hbin5gzohv5ftcospo6v4izvleqfb2uzygkbc7dok3yncuq", "uri": "at://did:plc:x2a3inabvfsn4wntrlbbndrv/app.bsky.feed.post/3liuq3jtnxk2e" }, "parent": { "cid": "bafyreigmk2kaj56ubcjeamp4uvbysvdq5uac63gf3i67h7i6sbkaskyjpe", "uri": "at://did:plc:x2a3inabvfsn4wntrlbbndrv/app.bsky.feed.post/3liuq3mrwks2e" } }, "createdAt": "2025-02-23T21:00:20.784Z" } }