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
ZyphraAI
15 atoms · 跨 4 天 · 首见 2026-05-06 · 最近 2026-06-25
三色: 🟦 fact 11 · 🟥 take 4 · stance ▲2/▼0/◆12
来源: X 15
时态: fresh:13 · aging:2
标签: 好信源:8 · 好观点:5 · 好思考:3 · 好数字:2 · 好问题:1
叙事 (narrative) (3)
- 2026-06-25 · 持续学习研究被评估为值得注意
feels like the most noteworthy paper about continual learning in quite some time
tags: 好观点 · → daily
- 2026-05-06
[老化中] · 实验计算成本低
well to be fair it's a 700M active model, one can afford quite a lot of experiments in this regime, speaking purely about compute
tags: 好观点 · → daily
value: qty=700M active model
- 2026-05-06
[老化中] · 研发与基础设施成本不低
but the research and infrastructure behind it are in no way cheap
tags: 好观点 · → daily
事实 (fact) (12)
- 2026-06-24 · found_plasticity_loss_scaling_law
We fit a scaling law for when plasticity loss sets in: T ∝ P^0.83
tags: 好观点·好思考 · → daily
- 2026-06-24 · plasticity_loss_scaling_relation
Onset grows sublinearly as a power-law with parameter count. So scaling delays the problem with sharply diminishing returns, but scale alone cannot save us from plasticity loss.
tags: 好观点·好思考 · → daily
- 2026-06-24 · studied_continual_learning_for_LLM
Zyphra is sharing our first work in continual learning where we study: Can LLMs learn forever from new data?
tags: 好思考·好问题 · → daily
- 2026-05-06 · ZAYA1-8B活跃参数量
With <1B active params, it outperforms open-weight models many times its size on math and reasoning, closing in on DeepSeek-V3.2 and GPT-5-High with test-time compute.
tags: 好数字·好信源 · → daily
value: qty=<1B
- 2026-05-06 · 推出 ZAYA1-8B 模型
Today we're releasing ZAYA1-8B, a reasoning MoE trained on @AMD and optimized for intelligence density.
tags: 好信源 · → daily
value: date=2026-05-06
- 2026-05-06 · ZAYA1-8B 发布
Today we're releasing ZAYA1-8B, a reasoning MoE trained on @AMD and optimized for intelligence density. With <1B active params, it outperforms open-weight models many times its size on math and re
tags: 好信源 · → daily
- 2026-05-06 · ZAYA1-8B 性能
With <1B active params, it outperforms open-weight models many times its size on math and re
tags: 好数字 · → daily
value: qty=<1B active params · metric=超过更大开源模型在数学及推理
- 2026-06-25 · 发布了关于持续学习的首个研究成果
Zyphra is sharing our first work in continual learning where we study: Can LLMs learn forever from new data?
tags: 好信源 · → daily
- 2026-05-31 · 发布ZAYA1-8B
@ZyphraAI launches ZAYA1-8B, the first MoE model trained on AMD Instinct MI300 hardware.
tags: 好信源 · → daily
- 2026-05-06 · 发布ZAYA1-8B模型
Today we're releasing ZAYA1-8B, a reasoning MoE trained on @AMD and optimized for intelligence density.
tags: 好信源 · → daily
- 2026-05-06 · ZAYA1-8B使用AMD训练
trained on @AMD
tags: 好信源 · → daily
- 2026-05-06 · ZAYA1-8B开源协议
We are releasing ZAYA1-8B open-weights under Apache 2.0
tags: 好信源 · → daily
时间轴 (近 20)
- 2026-06-25 ·
fact · 发布了关于持续学习的首个研究成果 · → daily
- 2026-06-25 ·
narrative · 持续学习研究被评估为值得注意 · → daily
- 2026-06-24 ·
fact · found_plasticity_loss_scaling_law · → daily
- 2026-06-24 ·
fact · plasticity_loss_scaling_relation · → daily
- 2026-06-24 ·
fact · studied_continual_learning_for_LLM · → daily
- 2026-05-31 ·
fact · 发布ZAYA1-8B · → daily
- 2026-05-06 ·
fact · 发布ZAYA1-8B模型 · → daily
- 2026-05-06 ·
fact · ZAYA1-8B使用AMD训练 · → daily
- 2026-05-06 ·
fact · ZAYA1-8B活跃参数量 · → daily
- 2026-05-06 ·
fact · ZAYA1-8B开源协议 · → daily
- 2026-05-06 ·
fact · 推出 ZAYA1-8B 模型 · → daily
- 2026-05-06 ·
fact · ZAYA1-8B 发布 · → daily
- 2026-05-06 ·
fact · ZAYA1-8B 性能 · → daily
- 2026-05-06 ·
narrative · 实验计算成本低 · → daily
- 2026-05-06 ·
narrative · 研发与基础设施成本不低 · → daily
← 实体目录 · 系统日志