Sequoia Capital

Kalshi's American Power Index (KPOW) transforms political power shifts into a single trackable number using prediction markets, combining current governance with future expectations for more objective political dynamics than traditional media or polls.

AlphaZero通过自我学习算法在数小时内超越需两年人类数据训练的AlphaGo，揭示了自主学习对AI突破的关键价值。

Mecka正在构建物理AI的数据和部署层，解决数据收集、硬件开发和全球运营等挑战。

TypeSafe的Jev模型通过实时决策改变软件自动化，24小时内成为Vercel最快采用的模型。

Databricks CEO用2天完成团队9个月的开发任务，AI技术使连接器开发周期从9个月压缩至7.5个月。

Peregrine公司通过整合现有数据而非收集新数据，颠覆传统数据驱动的商业模式。

中国开源模型正成为西方AI公司的重要依赖，但这也带来结构性风险。Qwen市场份额激增，蒸馏技术加剧中美模型差距。

Corma通过训练基础模型和强化学习应对AI攻击，填补防御性网络安全缺口，测试显示78%的防御失败率。

企业应构建自主AI能力以控制核心智能层，开源模型与后训练技术栈已具备商业可行性。

AI发展应聚焦算法和计算规模，而非依赖人类知识，这是Rich Sutton提出的'苦涩教训'核心观点。

Rich Sutton认为AGI不会是单一思维，而是多个系统协同学习，因‘大世界假设’指出单一系统无法掌握无限知识。

Preview平台通过AI原生工具解决视频创作行业痛点，已获超100家工作室采用。

小规模测试无法验证AI架构的真实潜力，强化学习等复杂任务需要基线级算力才能显现模型能力。

中国开源模型已占据西方AI公司训练栈核心位置，蒸馏技术加剧中美模型能力差距，西方需独立构建或接受滞后学习。

Bunkerhill Health的Carebricks平台通过AI代理整合医疗系统数据，使临床决策效率提升200%，并成功在UTMB Health部署20+智能代理。

Etched通过低电压推理和集群内存技术重构AI推理市场，其生产就绪的定制芯片将于2026年交付。

未来12-24个月90%的AI token将用于异步场景，推动自主代理技术发展，当前仍处于Copilot模式。

Cursor achieves 1TB model cross-continental synchronization during training by leveraging weight change patterns in RL, reducing transmission volume by 20x and ensuring model consistency.

The common trait among top founders is intense focus: they celebrate a win for a day and then immediately get back to the work. They are less distracted and don’t look around at what others are doing; they “mute the world” and build their own path. The UFC’s Dana White is a missionary-founder example, scaling from a tiny, no-money operation in Vegas to a global combat organization.

Suno transforms 90% of users from music consumers into creators. CEO Mikey Schulman's core insight is that the creation process itself is entertainment. Creating brings fulfillment that's in human DNA, and technology should let everyone experience this joy rather than just consume content.

Andrej Karpathy proposes a paradigm shift: neural networks as the primary computational substrate, with CPUs relegated to co-processors — redefining hardware/software hierarchy in AI-native systems.

Andrej Karpathy提出Software 3.0范式：软件开发正从手写代码（S1.0）和机器学习建模（S2.0）转向以提示工程驱动智能体协作的新阶段，核心能力是设计提示、调试行为与编排AI工作流。

Catalyst通过AI代理将投资想法转化为交易，降低用户门槛，适合对冲基金和散户投资者。

Peregrine的冷案代理能在1小时内处理300GB数据，帮助警方快速分析案件并定位嫌疑人。

自动化实验室通过增强人类代理而非取代人类，实现研究效率的指数级提升。

Sequoia Capital投资Sable，其AI员工Aidan具备实时浏览器交互能力，已获150+公司等待名单。

Anthropic提出AI代理需分层架构：知识层、执行层和协调层，强调tokens非可替代性需差异化角色分配。

Online RL is not the core for building models but an optimization method to enhance user experience after the model reaches a certain quality.

In large-scale RL training, numerical mismatches arise due to model version drift and floating-point precision differences, causing inconsistent log probabilities during inference and introducing training bias.

企业AI的难点在于管理和配置工具及技能，而非推理能力。