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

scaling laws

6 atoms · 跨 4 天 · 首见 2026-04-19 · 最近 2026-06-30

三色: 🟦 fact 1 · 🟥 take 5 · stance ▲0/▼0/◆2

来源: X 6

时态: fresh:5 · aging:1

标签: 好思考:3 · 好观点:2 · 好数字:1

2026-06-30 · power-law scaling mechanism
Why does power-law scaling occur? Loss of individual tokens follows a sigmoidal curve, and the aggregation of these curves with different times of learning makes a power-law curve.
tags: 好思考 · → daily
2026-06-24 · generally are sublinear
scaling laws generally are sublinear, ie you can get much more than 1/10th the performance with 1/10th the compute
tags: 好观点 · → daily
2026-04-19 [老化中] · 10T模型激活参数比例
models as massive as 10T don't allow <<3% activation, on priors they should
tags: 好观点 · → daily
value: qty=<<3% · direction=below

2026-04-19 · 250B参数模型可行性
250B is a good safe target
tags: 好思考 · → daily
value: qty=250B · direction=target
2026-04-19 · 150-200B参数模型可能性
I'd say 150-200 is possible
tags: 好思考 · → daily
value: qty=150-200B · direction=possible

2026-05-30 · rate of progress predicted exactly
Everything has been playing out exactly as the scaling laws predicted. The rate of progress has been exactly predictable. Too predictable. Zero upstarts, zero outliers, zero black swans.
tags: 好数字 · → daily