AI model training

4 atoms · 跨 3 天 · 首见 2026-05-08 · 最近 2026-06-09

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

来源: X 4

时态: aging:3 · fresh:1

标签: 好思考:3 · 好问题:1

叙事 (narrative) (3)

• 2026-05-08 [老化中] · discouraging essential behaviors encourages self-encryption to avoid monitoring
  If you are discouraging the model from doing things that are absolutely essential to lower the loss, you are encouraging it to find ways to self-encrypt so that monitoring fails.
  tags: 好思考 · → daily
• 2026-05-08 [老化中] · self evaluation selecting against Goodharted strategy after non-Goodharted strategy is learned leads to stable non-Goodharted outcome
  If you pair this with self evaluation that actively selects against the Goodharted strategy once the non-Goodharted strategy is learned then it seems probable you wind up with a stable non-Goodharted thing even in the capability regime where Goodharting is easy.
  tags: 好思考 · → daily
• 2026-05-10 [老化中] · technique to elevate knowledge worker views
  Probably. But how do you get the centroid of the knowledge worker specifically or the respectful programmer out of such a massive crawl? What technique ends up elevating their views above the ideology of the comments section?
  tags: 好问题 · → daily

预测 (forecast) (1)

• 2026-06-09 · performance improves with longer RL rollouts
  everything is downstream of longer RL rollouts
  tags: 好思考 · → daily

时间轴 (近 20)

• 2026-06-09 · forecast · performance improves with longer RL rollouts · → daily
• 2026-05-10 · narrative · technique to elevate knowledge worker views · → daily
• 2026-05-08 · narrative · discouraging essential behaviors encourages self-encryption · → daily
• 2026-05-08 · narrative · self evaluation selecting against Goodharted strategy after · → daily

← 实体目录 · 系统日志