kalomaze

6 atoms · 跨 4 天 · 首见 2026-04-16 · 最近 2026-06-14

三色: 🟦 fact 0 · 🟥 take 6 · stance ▲1/▼1/◆4

来源: X 6

时态: fresh:4 · aging:2

标签: 好观点:5 · 好思考:3 · 好问题:1

叙事 (narrative) (6)

• 2026-04-16 ⭐ [老化中] · 认为合成数据生产可成为形式化外循环优化问题
  i am saying that there is a path to synthetic data production as a formalized outer loop optimization problem, where some generator produces data augmentations that are evaluated wrt "is this better in a distributed, consistent way for held out prediction across the manifold"
  tags: 好观点·好思考·好问题 · → daily
• 2026-04-16 [老化中] · 认为最佳训练数据不是自然网络文本
  something that almost nobody internalizes is that the best data to train a language model on to reduce prediction error on natural webtext is almost certainly not the natural webtext itself
  tags: 好观点·好思考 · → daily
• 2026-06-14 · stated that observed monotonic-ish improvements in accuracy on API reasoning with varying reasoning parameter versus routed to different pretrains
  i've observed monotonic-ish improvements on accuracy in reasoning on the api with varying the reasoning parameter and it being routed to different pretrains would be highly inconsistent with basically ~all statistical evidence i have
  tags: 好思考 · → daily
• 2026-06-14 · stated that things that newer opuses seem to fail at much, much more is "not paraphrasing shit that is supposed to be identical"
  the thing that newer opuses seem to fail at much, much more is "not paraphrasing shit that is supposed to be identical"
  tags: 好观点 · → daily
• 2026-06-13 · stated working on closest thing to a true "big model smell" eval
  i am trying to work on the closest thing possible to a true "big model smell" eval
  tags: 好观点 · → daily
• 2026-05-29 · 评论
  ngl they spent actual time on the new classifiers and they seem pretty aggro wrt anything that looks even adjacent to cybersec
  tags: 好观点 · → daily
  value: direction=neutral

时间轴 (近 20)

• 2026-06-14 · narrative · stated that things that newer opuses seem to fail at much, m · → daily
• 2026-06-14 · narrative · stated that observed monotonic-ish improvements in accuracy · → daily
• 2026-06-13 · narrative · stated working on closest thing to a true "big model smell" · → daily
• 2026-05-29 · narrative · 评论 · → daily
• 2026-04-16 · narrative · 认为最佳训练数据不是自然网络文本 · → daily
• 2026-04-16 · narrative · 认为合成数据生产可成为形式化外循环优化问题 · → daily

← 实体目录 · 系统日志