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

jsuarez

6 atoms · 跨 2 天 · 首见 2026-06-21 · 最近 2026-06-22

三色: 🟦 fact 6 · 🟥 take 0 · stance ▲2/▼0/◆2

来源: X 6

时态: fresh:6

标签: 好数字:2 · 好观点:2 · 好信源:1 · 好思考:1

2026-06-21 · experiment turnaround speed claim
That gives us a quick experiment turnaround.
tags: 好思考 · → daily
value: qty=quick experiment turnaround
2026-06-22 · can train on new game faster than LLM fine-tune
We can train it on the 11th game in less time than it takes to fine tune an LLM on the 11th
tags: 好观点 · → daily
2026-06-21 · deep learning research tenure
I've been in deep learning research for over a decade.
tags: 好信源 · → daily
value: qty=over a decade
2026-06-21 · training speed claim: 50k param model
We use a 50k param model because we can crank 10M samples per second on it (20 once optimized).
tags: 好数字 · → daily
value: qty=10M samples per second
2026-06-21 · general method solves entire problem class
The method I am working on is fully general in that it solves an entire class of problems that we cannot otherwise solve with existing methods.
tags: 好观点 · → daily
2026-06-21 · training comparison with alternatives
Maybe we 'only' train many client problems at 5M steps/second, but the most common alternatives train 5k steps/second. I routinely see people working in the 100s of steps/second. So in order to solve the same problem that we solve on 1 GPU, you need a 1000 GPU setup, and you don't have infra for that.
tags: 好数字 · → daily
value: qty=1000x speedup