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
Parallax
12 atoms · 跨 3 天 · 首见 2026-05-29 · 最近 2026-06-02
三色: 🟦 fact 4 · 🟥 take 8 · stance ▲6/▼0/◆6
来源: X 12
时态: fresh:12
标签: 好观点:7 · 好思考:7 · 好信源:1
叙事 (narrative) (9)
- 2026-05-29 · perplexity
Parallax has better perplexity
tags: 好观点·好思考 · → daily
value: qty=better
- 2026-05-29 · downstream accuracy
Parallax has better downstream accuracy
tags: 好观点·好思考 · → daily
value: qty=better
- 2026-05-29 · decode kernel performance
a decode kernel that matches or beats FlashAttention
tags: 好观点·好思考 · → daily
value: direction=matches or beats FlashAttention
- 2026-06-02 · modification transferability
The cleanest modifications often transfer best; records built on heavy tuning transfer less reliably.
tags: 好思考 · → daily
- 2026-06-01 · not linear attention or SSM
Parallax and LLA are not linear complexity attention.
tags: 好思考 · → daily
- 2026-06-01 · compatible with contextual sparsity methods
_Parallax is compatible with contextual sparsity methods such as sliding window, block-sparse, and the most recent MSA by @MiniMax_AI._
tags: 好思考 · → daily
- 2026-06-01 · standard Attention with KV cache
Structurally they're like standard Attention which keeps a KV cache for inference.
tags: 好思考 · → daily
- 2026-06-02 · compared to Softmax Attention
Parallax is uniformly stronger than Softmax Attention across all records.
tags: 好观点 · → daily
- 2026-06-02 · optimizer transferability
Optimizers don't transfer to Parallax with the same magnitude, which confirms the optimizer–architecture interaction from the Parallax paper.
tags: 好观点 · → daily
预测 (forecast) (1)
- 2026-06-02 · improvement room
These are preliminary results, I believe both the Parallax architecture and the optimizer side have room to improve.
tags: 好观点 · → daily
事实 (fact) (2)
- 2026-06-01 · already supported sliding window on GitHub
We already supported sliding window in the github.
tags: 好信源 · → daily
- 2026-05-29 · requires Muon optimizer to move beyond Softmax Attention
Muon optimizer is crucial for Parallax to move beyond Softmax Attention
tags: 好观点 · → daily
时间轴 (近 20)
- 2026-06-02 ·
narrative · compared to Softmax Attention · → daily
- 2026-06-02 ·
narrative · optimizer transferability · → daily
- 2026-06-02 ·
narrative · modification transferability · → daily
- 2026-06-02 ·
forecast · improvement room · → daily
- 2026-06-01 ·
narrative · not linear attention or SSM · → daily
- 2026-06-01 ·
narrative · compatible with contextual sparsity methods · → daily
- 2026-06-01 ·
narrative · standard Attention with KV cache · → daily
- 2026-06-01 ·
fact · already supported sliding window on GitHub · → daily
- 2026-05-29 ·
narrative · perplexity · → daily
- 2026-05-29 ·
narrative · downstream accuracy · → daily
- 2026-05-29 ·
narrative · decode kernel performance · → daily
- 2026-05-29 ·
fact · requires Muon optimizer to move beyond Softmax Attention · → daily
← 实体目录 · 系统日志