The AI hype cycle just met the income statement: two design-software leaders sent a message most AI vendors dread — popularity doesn't pay the inference bill. Canva slowed, Figma eats inference costs.
Canva 主动叫停 AI 功能铺开、Figma 自吞推理成本——两家头部设计公司的财报把 AI 转型最隐秘的代价摆上台面：受欢迎程度不等于健康的单位经济模型。
Most AI product testing treats the user as one interchangeable blob. MatrAIx bets the opposite: sample a persona — a 1,290-dimension profile — and build a user simulator that behaves like real users.
MatrAIx 实战：用 1290 维人格画像把用户实例化成 LLM 智能体，在问卷、聊天、网页、原生 App 里跑可复现评测。哈佛/MIT 领衔 200+ 科学家参与，100 万条质控人格数据今日可上手。
Claude Code 自动模式即将默认开启：每次工具调用前过风险分类器，Hard deny 永不自动放行。5 天生效前先改好权限配置——切换前要检查这几处。
A 25-year gap, closed in seven days. Microsoft Research's Dimitris Papailiopoulos, working with GPT-5.6 and Claude Fable 5, produced a proof that a simple two-step algorithm solves MIMO detection.
微软研究院首席研究员与 GPT-5.6、Claude Fable 5 合作，七天证明多项式时间算法能在最大似然阈值处精确恢复全部比特——拿下悬置 25 年的 MIMO 检测难题。
Monday, August 10, 2026 — Inference costs are the story: DeepSeek signals a big API price hike, Amazon reportedly burns money running Claude at scale, and memory-chip shortages keep squeezing supply.
2026年8月10日 AI 早报：180 万美元连亚马逊都烧不起 Claude、GPT-5.6 解 25 年数学难题、中国团队跑通数据层 RSI 等今日亮点 + 7 篇全球重要新闻。
For two years the story of AI progress has been compute. BigBang-V1's open-source release points at a quieter bottleneck: data. 100% AI-synthesized training data may run the data layer's RSI loop.
当人类出题速度追不上模型：35B 开源模型 BigBang-V1 后训练数据 100% AI 自主合成（出题、解题、验证全交给 AI），8 项基准拿遍 35B 档最高分，甚至超过 1.6T 的 DeepSeek V4 Pro。
On August 9, OpenAI switched off Atlas, its AI-native browser — 292 days after launch. The most decisive of four retreats: AI browsers were a feature, not a destination.