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
MAI
30 atoms · 跨 3 天 · 首见 2026-06-02 · 最近 2026-06-22
三色: 🟦 fact 19 · 🟥 take 11 · stance ▲4/▼3/◆19
叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom
状态只算低频策展源 (群/卖方); X firehose 仅作背景音量
来源: X 30
时态: fresh:30
标签: 好数字:15 · 好思考:8 · 好观点:6 · 好信源:1
🟨 AI 综合 · junior analyst 概览
展开 AI 综合 (灰色 · 非市场结论 · 点击数字溯源到原 atom)
model: deepseek-chat · 2026-07-12 · 默认折叠
MAI 近期发布了模型，并以 8K B200 集群训练 1T-A35B 架构
¹¹，且将稀疏度扩展推至更远
¹。但优化后 MFU 仅为 20%，令人意外
¹；其截断策略可能限制最终性能
¹；自主智能体表现也存在显著差距
¹。融资层面，MAI 已筹 2500 万美元
¹，并采用自研爬虫而非 Common Crawl
¹。在 McKinsey 任务上，MAI 以低于 GPT-5.5 10 倍的成本实现更高胜率
¹¹，且开源模型研究被视作推动平台生态的手段而非核心
¹。最终 RL 阶段批次大小近 10 亿 token，此前短上下文阶段批次也始终不低于预训练批次
¹¹。
🟦 客观事实 (facts) (19)
- <a id="atom-3d6dbd0ba7297376"></a>🟦 2026-06-22
X·@yuchen_wu · 融资金额 · 融资_
we’ve raised $25M and built MAI to fix that
tags: 好数字 · → daily
value: qty=$25M
- <a id="atom-49d591b5c0d8c34a"></a>🟦 2026-06-04
X·@Grad62304977 · final RL stage batch size · 技术路线
MAI used a batch size of almost 1B tokens during their final RL stage
tags: 好数字 · → daily
value: qty=~1B tokens
📷 原图
- <a id="atom-6c5bf01af18c30ae"></a>🟦 2026-06-04
X·@Grad62304977 · shorter context stage batch size · 技术路线
even in their previous shorter context stages, the batch sizes were always bigger or equal to the pretraining batch size
tags: 好数字 · → daily
- <a id="atom-42900fa7abf642d6"></a>🟦 2026-06-04
X·@Laz4rz · 最近两天发布了模型 · 模型发布
Last two days we got Gemma, MAI and Nemotron Ultra
tags: 好数字 · → daily
- <a id="atom-317058ca77a34e48"></a>🟦 2026-06-02
X·@mustafasuleyman · cost vs GPT-5.5 on McKinsey's tasks · 定价效率
MAI delivered the highest win rate, outperforming GPT-5.5 on quality, while being 10x lower on cost.
tags: 好数字 · → daily
value: qty=10x lower
📷 原图
- <a id="atom-5936ee886d464193"></a>🟦 2026-06-02
X·@mustafasuleyman · win rate vs GPT-5.5 on McKinsey's tasks · 竞争排名
MAI delivered the highest win rate, outperforming GPT-5.5 on quality
tags: 好数字 · → daily
value: qty=highest
📷 原图
- <a id="atom-9353d64ddccc4f8a"></a>🟦 2026-06-02
X·@stochasticchasm · 训练集群配置 · 数据中心算力
8K B200s for a 1T-A35B seems like quite a comfortable decision
tags: 好数字 · → daily
value: qty=8K B200s · date=na
- <a id="atom-9a7bfec78aaae411"></a>🟦 2026-06-02
X·@stochasticchasm · 数据爬取方式 · 技术路线
they even skip CC and go to a proprietary crawler
tags: 好数字 · → daily
value: qty=na · date=na
- <a id="atom-04519935b5a20acd"></a>🟦 2026-06-02
X·@stochasticchasm · 无合成数据策略 · 技术路线
we don't expose the model to synthetic data in mid-training either!
tags: 好数字 · → daily
value: qty=na · date=na
📷 原图
- <a id="atom-87e7555b3e1e18f3"></a>🟦 2026-06-02
X·@stochasticchasm · 正则化使用 · 技术路线
dropout 0.15 is quite large
tags: 好数字 · → daily
value: qty=0.15 · date=na
- <a id="atom-8b79062f4cf20325"></a>🟦 2026-06-02
X·@stochasticchasm · 采样策略 · 技术路线
they use top-p as a logit filter before doing the sampling
tags: 好数字 · → daily
value: qty=na · date=na
- <a id="atom-49ef22b5efbc2f66"></a>🟦 2026-06-02
X·@stochasticchasm · 自蒸馏中dropout · 技术路线
dropout is back for self-distillation as well
tags: 好数字 · → daily
value: qty=na · date=na
- <a id="atom-46f4e1409530db6b"></a>🟦 2026-06-02
X·@stochasticchasm · goodput数据 · 盈利能力
really nice to see goodput numbers
tags: 好数字 · → daily
value: qty=na · date=na
- <a id="atom-3e82310d00a8774e"></a>🟦 2026-06-02
X·@stochasticchasm · 缩放配置 · 技术路线
very nice of them to share scaling ladder configurations
tags: 好信源 · → daily
value: qty=na · date=na
- <a id="atom-fa684e1e4e72ba80"></a>🟦 2026-06-02
X·@stochasticchasm · 负载均衡策略 · 技术路线
softmax routing + global-batch LBL makes this the qwen-flavored load balancing style instead of deepseek-flavored
tags: 好思考 · → daily
value: qty=na · date=na
- <a id="atom-d3178edc5b6fc10f"></a>🟦 2026-06-02
X·@stochasticchasm · SFT数据来源 · 技术路线
they just passively collect traces from RL for SFT
tags: 好思考 · → daily
value: qty=na · date=na
- <a id="atom-f51d82cc77a64da3"></a>🟦 2026-06-02
X·@stochasticchasm · RL环境设计 · 技术路线
they're using both BT reward modeling as well as LLM judges from preference and safety tuning
tags: 好思考 · → daily
value: qty=na · date=na
- <a id="atom-4a23fbeab75536e5"></a>🟦 2026-06-02
X·@stochasticchasm · 模型架构选择 · 技术路线
architecture seems pretty safe/solid. nothing incredibly risky here. QK-norm, interleaved local/global, 64x sparsity, sandwich norm, GQA-8, head dim 128
tags: 好观点 · → daily
value: qty=na · date=na
📷 原图
- <a id="atom-dfd448fd6ef39ddd"></a>🟦 2026-06-02
X·@stochasticchasm · GRPO公式 · 技术路线
pretty vanilla GRPO formulation
tags: 好观点 · → daily
value: qty=na · date=na
🟥 多头 takes (bullish) (2)
- <a id="atom-c0ced89b83daf7dc"></a>🟥 2026-06-04
X·@willccbb · 开源模型研究对其核心业务不太重要，但能推动平台/生态系统完善 · 平台战略
slightly different situation for nvidia / google / MAI where the open model research is less central to their core business but drives platform/ecosystem completeness
tags: 好思考 · → daily
- <a id="atom-3ae0aff30f92d4a8"></a>🟥 2026-06-02
X·@stochasticchasm · 稀疏度扩展 · 技术路线
good confirmation that sparsity can be pushed much further, even by an efficiency gain metric
tags: 好观点 · → daily
value: qty=na · date=na
🟥 空头 takes (bearish) (3)
- <a id="atom-ca892fc1c415967a"></a>🟥 2026-06-02
X·@stochasticchasm · MFU表现 · 盈利能力
i'm genuinely a bit surprised that they hit only 20% MFU after all this optimization
tags: 好数字 · → daily
value: qty=20% · date=na
- <a id="atom-c26ea29646346cca"></a>🟥 2026-06-02
X·@stochasticchasm · agent性能 · 技术路线
sheer agentic performance deltas here
tags: 好数字 · → daily
value: qty=na · date=na
- <a id="atom-ec41cbf050bacd4d"></a>🟥 2026-06-02
X·@stochasticchasm · 截断策略 · 技术路线
nooo guys nooo this can cap final performance nooo
tags: 好思考 · → daily
value: qty=na · date=na
🟥 中性 takes (neutral) (6)
- <a id="atom-57209251c283a85d"></a>🟥 2026-06-02
X·@stochasticchasm · 数据预处理工具 · 技术路线
trafilatura for HTML -> text is a very safe choice
tags: 好思考 · → daily
value: qty=na · date=na
- <a id="atom-493b41e9e3d31bf2"></a>🟥 2026-06-02
X·@stochasticchasm · RL训练策略 · 技术路线
only hard ratio clipping and no token or sequence masking strategies is surprising
tags: 好思考 · → daily
value: qty=na · date=na
- <a id="atom-2cdbaa167dc023bf"></a>🟥 2026-06-02
X·@stochasticchasm · 优化器选择 · 技术路线
muon's been pretty well-adopted by now, so i'm a bit surprised at this
tags: 好思考 · → daily
value: qty=na · date=na
- <a id="atom-354bdbc42c991ef1"></a>🟥 2026-06-02
X·@stochasticchasm · RL自蒸馏策略 · 技术路线
the most interesting aspect is this self-distillation strategy that they use to resume runs after mysterious collapses
tags: 好观点 · → daily
value: qty=na · date=na
- <a id="atom-d6817256eb97c1a5"></a>🟥 2026-06-02
X·@stochasticchasm · SFT数据位置 · 技术路线
everyone is moving SFT data further back into the pipeline for MoEs
tags: 好观点 · → daily
value: qty=na · date=na
- <a id="atom-dcf071b0f1e5dada"></a>🟥 2026-06-02
X·@stochasticchasm · 门控奖励信号 · 技术路线
first tech report to explicitly mention gated reward signals being important
tags: 好观点 · → daily
value: qty=na · date=na
⏱ 时间轴 (近 20)
- 🟦 2026-06-22 ·
fact · 融资金额 · → daily
- 🟦 2026-06-04 ·
fact · final RL stage batch size · → daily
- 🟦 2026-06-04 ·
fact · shorter context stage batch size · → daily
- 🟦 2026-06-04 ·
fact · 最近两天发布了模型 · → daily
- 🟥 2026-06-04 ·
narrative · 开源模型研究对其核心业务不太重要，但能推动平台/生态系统完善 · → daily
- 🟦 2026-06-02 ·
fact · cost vs GPT-5.5 on McKinsey's tasks · → daily
- 🟦 2026-06-02 ·
fact · win rate vs GPT-5.5 on McKinsey's tasks · → daily
- 🟦 2026-06-02 ·
fact · 训练集群配置 · → daily
- 🟦 2026-06-02 ·
fact · 模型架构选择 · → daily
- 🟦 2026-06-02 ·
fact · 负载均衡策略 · → daily
- 🟦 2026-06-02 ·
fact · 缩放配置 · → daily
- 🟥 2026-06-02 ·
fact · 稀疏度扩展 · → daily
- 🟥 2026-06-02 ·
fact · 数据预处理工具 · → daily
- 🟦 2026-06-02 ·
fact · 数据爬取方式 · → daily
- 🟦 2026-06-02 ·
fact · 无合成数据策略 · → daily
- 🟥 2026-06-02 ·
fact · MFU表现 · → daily
- 🟦 2026-06-02 ·
fact · GRPO公式 · → daily
- 🟥 2026-06-02 ·
fact · RL训练策略 · → daily
- 🟥 2026-06-02 ·
fact · 优化器选择 · → daily
- 🟦 2026-06-02 ·
fact · 正则化使用 · → daily
← 实体目录 · 系统日志
