以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
RL
13 atoms · 跨 10 天 · 首见 2026-04-27 · 最近 2026-06-28
三色: 🟦 fact 2 · 🟥 take 11 · stance ▲6/▼3/◆4
来源: X 13
时态: fresh:8 · aging:4 · stale:1
标签: 好观点:6 · 好思考:5 · 好信源:3 · 好数字:2
叙事 (narrative) (9)
- 2026-04-28
[老化中] · compute scaling working on any domain
we take this kind of plot for granted now, but RL compute scaling working on literally any domain (you can verify?) is still beautiful to me
tags: 好观点·好思考 · → daily
- 2026-06-28 · 应用时机
people are asking too much of RL. Should only come once whole agentic system is modeled
tags: 好思考 · → daily
value: direction=after agentic system modeled
- 2026-06-04 · 与蒸馏趋同于策略梯度
Self-Distilled Policy Gradient. Token-level rewards. Credit assignment. Self-distillation. RL and distillation are converging toward the same idea: Policy gradients. It always has been. It always will be.
tags: 好思考 · → daily
- 2026-05-10
[老化中] · 局限
If there is heuristic for solving the game simple enough to be coded by codex, RL will find it and optimize towards it. But it is clear that such heuristic is not generally intelligent and will not generalise.
tags: 好思考 · → daily
- 2026-05-10
[老化中] · 瓶颈
The bottleneck to RL generalization lies in environments where simple heuristics get you far and it’s easy to get stuck in local optimum.
tags: 好思考 · → daily
- 2026-05-29 · 技术优化定位
RL is one of the optimizations
tags: 好观点 · → daily
- 2026-05-21 · 零售商分类
Yeah winners and losers in retail space. You have winners like WSM RL etc.
tags: 好观点 · → daily
value: direction=winner
- 2026-05-14 · 品牌持久且当前超受欢迎
$RL, best to ever do it... That one is kind of everlasting, but definitely has ups and downs. Now it’s super popular again
tags: 好观点 · → daily
value: date=2026-05-14
- 2026-04-27
[老化中] · 最优训练方式
RL'ing on everything together is just the way
tags: 好观点 · → daily
反驳 (refute) (1)
- 2026-05-07 · 对比遗忘研究论文推荐
read this paper for reference on why RL forgets less compared to SFT
tags: 好信源 · → daily
立场 (position) (1)
展开 1 条老旧 / 已过期
- 2026-05-09
[老旧] · 标签
This is unfortunately mostly a bearish take on RL
tags: 好观点 · → daily
事实 (fact) (2)
- 2026-05-21 · 美国同店销售
US Retailers who have reported so far like $RL $WSM and $TGT have seen strongly accelerating US comparable sales last quarter.
tags: 好数字·好信源 · → daily
value: direction=strongly accelerating
- 2026-05-21 · 营收增长
Revenue growth overall has strongly accelerated due to strong international growth & online boom in many cases.
tags: 好数字·好信源 · → daily
value: direction=strongly accelerated
时间轴 (近 20)
- 2026-06-28 ·
narrative · 应用时机 · → daily
- 2026-06-04 ·
narrative · 与蒸馏趋同于策略梯度 · → daily
- 2026-05-29 ·
narrative · 技术优化定位 · → daily
- 2026-05-21 ·
fact · 美国同店销售 · → daily
- 2026-05-21 ·
fact · 营收增长 · → daily
- 2026-05-21 ·
narrative · 零售商分类 · → daily
- 2026-05-14 ·
narrative · 品牌持久且当前超受欢迎 · → daily
- 2026-05-10 ·
narrative · 局限 · → daily
- 2026-05-10 ·
narrative · 瓶颈 · → daily
- 2026-05-09 ·
position · 标签 · → daily
- 2026-05-07 ·
refute · 对比遗忘研究论文推荐 · → daily
- 2026-04-28 ·
narrative · compute scaling working on any domain · → daily
- 2026-04-27 ·
narrative · 最优训练方式 · → daily
← 实体目录 · 系统日志