a16z New Media团队成立一周年,提供“go-direct as a service“服务,帮助初创公司通过自制内容和直接叙事赢得关注。该团队具备四个核心运作方式:Launch Videos(自制发布视频)、Editorial(编辑写作支持)、Distribution(分发保证)和Portfolio Services(被投企业服务)。与传统机构不同,a16z拥有自有渠道可保证分发效果,且团队成员与创始人认知水平相当。核心策略是将创始人个人打造为高信号品牌,而非仅依赖公司品牌。
a16z 将其投资者关系(Investor Relations)部门更名为 Global Partnerships,标志着该风投公司的服务范围从传统IR扩展至全球合作伙伴关系。新部门旨在帮助创始团队获取最高层级的国家和机构资本、客户及分销渠道,并在进入新市场时具备地缘政治敏感度,以将前沿技术引入美国及其盟友。此举旨在把最优秀的创始人与所需资本和国家资源对接,使LP在公司国际化扩张中获益。
文章讨论当代世界中关键“抽象”(reliable interfaces built out of less-than-reliable parts)正在失效或泄漏的现象。作者认为AI末日论关注“将被建造的东西”,而主流末日论关注“将崩溃的东西”,但真正的洞见在于认识到许多支撑世界运转的抽象——如全球贸易、能源市场、美元体系——不会彻底崩溃,却会在地缘政治冲击下持续泄漏其底层复杂性。以电子商务结账、美国财政部纸质支票案例、霍尔木兹海峡危机导致的全球油价市场异动为例,论证当政治承诺或个人感知介入复杂系统时,反身性反馈会放大系统性风险。Mark Carney在达沃斯的发言被引用为对20世纪全球抽象(统一市场、统一规范)终结的官方确认。
旧金山科技社交场景已主要由“终端在线”人群驱动,社交图谱首先在X(Twitter)上建立,随后通过私人群聊和Partiful活动实现线下聚会。社交资本从线下会议和咖啡聊转向线上互相关注和内容互动,形成以“mutuals”为证明工作的新模式。许多人通过长期线上内容消费在X上相识并最终线下见面,甚至找到恋爱对象。新搬入的居民常只认识通过X或X用户聚集的活动认识的人。传统意义上场所提供的社交“工作”已被线上社交基础设施取代,活动报名通过Partiful排队和“门卫”控制。a16z的新媒体 fellowship(New Media Fellowship)采用X群聊和Partiful组织线下聚会,未使用名片或破冰环节。该文认为这些模式可能向其他城市扩散,标志着整个世代建立信任、寻找爱情和招聘方式的潜在转变。
本文指出,互联网上通过开放标准实现的开放与控制平衡曾保障了学习自由,但随着AI技术普及,出版商正通过付费墙、服务条款和法院禁令限制公开信息的访问,可能把学习自由从公共资源转变为受合同和技术壁垒约束的特权。此趋势将提高AI训练和使用成本,削弱用户对合法内容的分析和摘要能力,并因大公司更易适应复杂规则而导致市场集中度上升。历史判例(Ticketmaster案、KPMG链接争议、书籍首次销售 doctrine、X v. Bright Data)表明合同不能完全剥夺公众对公开数据的合理使用权。作者建议通过市场、技术与政策相结合的定制方案(如robots.txt标准演进)来在AI时代维护学习自由。
This article examines the constitutional limits on state authority to regulate artificial intelligence, focusing on the dormant Commerce Clause. It outlines three doctrinal components—anti‑discrimination, Pike balancing, and the anti‑extraterritorial principle—and illustrates their application through recent case law (Pike v. Bruce Church, National Pork Producers v. Ross, Minnesota insulin pricing case). The piece notes that, in 2024, states enacted over 100 AI‑related statutes while Congress passed none, and it highlights California AB 1018 as a bill that could impose hundreds of millions of dollars in compliance costs and effectively set de‑facto national standards, raising dormant Commerce Clause concerns. The analysis advises states to target harmful AI uses within their borders rather than extraterritorial model‑development regulation.
Consumer AI apps are rapidly adopting $200/month hyper‑premium subscription tiers that serve narrow, high‑intensity use cases rather than the mass market, enabling niche‑focused products with favorable unit economics and signaling a shift away from ad‑supported growth models toward deep, specialized offerings.
The third batch of Open Source AI Grant recipients was announced on 2025‑06‑26. The grant program provides non‑investment funding to hackers, researchers, and small teams developing AI outside major labs. This batch emphasizes evaluating frontier large language model (LLM) capabilities and testing their limits, highlighting benchmarks such as SWE‑Bench and ARC Prize that go beyond saturated metrics to assess genuine reasoning and real‑world problem solving. Experimental work by truth_terminal, Pliny, and Janus is cited for pushing AI boundaries, revealing unexpected abilities and raising questions about AI interaction with the real world. Additional grants target top open‑source AI infrastructure teams.
In early 2025 an anonymous document branded "Unpatched AI" published >100 previously unknown Microsoft Access and Office 365 vulnerabilities with full technical proofs, stack traces and exploit chains, indicating an autonomous LLM‑steered vulnerability‑research pipeline that combines fuzzing, symbolic execution and generative narration. The security community recognized the findings as genuine and highly automated, prompting the realization that autonomous systems are beginning to compete—and in some cases outperform—human researchers in offensive security tasks, moving up public bug‑bounty leaderboards and scaling attack‑surface coverage without human guidance. Traditional penetration testing relies on periodic, human‑driven engagements that cannot keep pace with rapid software change; the 2025 Verizon DBIR notes >67 % of breaches involve unpatched flaws >90 days old despite recent assessments. A new class of AI‑native pentesting platforms unbundles expert‑labor constraints by coupling LLMs with exploit tooling, real‑time telemetry and proprietary data, offering fully autonomous agents or copilot‑style assistance that execute exploits in safe sandboxes, verify findings, and generate actionable reports. Early platforms promised automation but suffered shallow coverage, static detection logic, poor cloud‑native support and alert fatigue ("50 000 criticals, zero real"). Current systems excel at low‑hanging issues (XSS, SSRF, misconfigs) and business‑logic flaws that can be inferred from intent, but still lag in complex chained authorizations, race conditions and environment‑specific contexts requiring deep contextual reasoning. Regulatory frameworks (SOC 2, PCI, ISO 27001) require human‑led assessments, creating auditability and liability gaps for fully autonomous tools. The emerging shift is toward continuous, AI‑augmented testing integrated into CI/CD pipelines, blurring the line between testing, pentesting and red‑teaming.
Based on a June 2025 survey of 100 enterprise CIOs across 15 industries, generative‑AI budgets are growing faster than expected (average ~75% YoY), with spend shifting from experimental innovation funds to core IT and business‑unit budgets. Multi‑model deployments are now standard (37% using 5+ models), driven by task‑specific model differentiation rather than just vendor‑lock‑in concerns. Model costs have dropped roughly ten‑fold per year, improving price‑performance for closed‑source small/medium models; as a result, fine‑tuning is declining in favor of prompt engineering and long‑context solutions. Reasoning models are in early testing but already showing strong adoption for OpenAI’s o3 (23% in production vs 3% for DeepSeek). Enterprises are moving toward direct procurement from model providers or via Databricks, seeking early access and lower switching costs—though agentic workflows increase lock‑in because extensive prompt tuning is model‑specific. External benchmarks (e.g., LM Arena) are now a key filter in procurement, complementing internal evaluation. The ecosystem of third‑party AI applications has matured, driving a buy‑vs‑build shift; over 90% of respondents are testing third‑party customer‑support apps, while regulated sectors like healthcare still favor in‑house development.
AI‑native startups are reaching revenue milestones far faster than pre‑AI benchmarks, with median enterprise ARR exceeding $2 M and median consumer ARR reaching $4.2 M within the first year, enabling Series A funding within 8–9 months after monetization.
The article argues that enterprise AI startups succeed by adopting implementation‑heavy, services‑oriented models similar to early Salesforce, ServiceNow and Workday, rather than pure product‑led growth. It contends that AI agents require deep integration, data ingestion and ongoing management, which creates a need for forward‑deployed engineers and professional services that lower early gross margins but build durable moats and higher margins later. The author claims that even model providers are hiring solutions engineers, and that controlling the data ingestion point is critical for building a system of work that rivals traditional systems of record.