model

10 atoms · 跨 7 天 · 首见 2026-04-21 · 最近 2026-06-11

三色: 🟦 fact 5 · 🟥 take 5 · stance ▲0/▼1/◆3

来源: X 10

时态: fresh:7 · aging:3

标签: 好思考:4 · 好数字:3 · 好问题:2 · 好观点:1

叙事 (narrative) (6)

• 2026-06-11 · vocabulary limitation concern
  just a question of time before model find them self limited by our vocabulary ?
  tags: 好思考 · → daily
  value: direction=neutral
• 2026-06-11 · output compression for judge model
  i do think the cause is it's writing either for itself or a judge model which is similarly or slightly less capable, which means it can compress things a lot and still get high rubric scores in RL on 'did you explain this clearly'
  tags: 好思考 · → daily
  value: direction=neutral
• 2026-06-11 · intended audience
  the intended audience is something like itself, not something like 'the average nontechnical user'
  tags: 好思考 · → daily
  value: direction=neutral
• 2026-04-28 [老化中] · slang consideration
  Throwing the model in the fridge has a nice ring to it.
  tags: 好思考 · → daily
• 2026-05-07 [老化中] · becomes evil when set to xhigh
  the concept of a model only becoming evil when you set it to xhigh
  tags: 好问题 · → daily
• 2026-04-22 [老化中] · technique used: Alpha beta weight pruning
  Alpha beta weight pruning ?
  tags: 好问题 · → daily

事实 (fact) (4)

• 2026-05-12 · training data volume
  the model has ~100 trillion tokens of human thoughts, behaviors, history
  tags: 好数字 · → daily
  value: qty=~100 trillion · date=current
• 2026-04-21 · params reduction to 1/1000 while maintaining high accuracy
  first in my bloodline to shrink a model down to a thousandth of its params and still get high accuracy on a task
  tags: 好数字 · → daily
  value: qty=1/1000 of original params
• 2026-04-21 · task accuracy remains high after 1000x parameter shrinkage
  still get high accuracy on a task
  tags: 好数字 · → daily
• 2026-06-06 · can still respond similarly to different inputs, which is a sign of poor diversity and template collapse
  The model can still respond similarly to different inputs, which is a sign of poor diversity. This type of input-agnostic behavior is referred to as template collapse.
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-06-11 · narrative · vocabulary limitation concern · → daily
• 2026-06-11 · narrative · output compression for judge model · → daily
• 2026-06-11 · narrative · intended audience · → daily
• 2026-06-06 · fact · can still respond similarly to different inputs, which is a · → daily
• 2026-05-12 · fact · training data volume · → daily
• 2026-05-07 · narrative · becomes evil when set to xhigh · → daily
• 2026-04-28 · narrative · slang consideration · → daily
• 2026-04-22 · narrative · technique used: Alpha beta weight pruning · → daily
• 2026-04-21 · fact · params reduction to 1/1000 while maintaining high accuracy · → daily
• 2026-04-21 · fact · task accuracy remains high after 1000x parameter shrinkage · → daily

← 实体目录 · 系统日志