Arena的创始人兼首席执行官Anastasios Anjapaloulos对飞速发展的AI行业，特别是在模型评估、地缘政治动态和新兴商业模式方面，提出了坦率而富有洞察力的观点。 他的公司Arena定位于AI的核心评估平台，超越静态基准测试，衡量AI在实际世界中的表现。它在模型由真人实际使用时，基于事实准确性、可控性、人类偏好和幻觉等因素进行评估。这种直接反馈有助于实验室改进模型，并向整个生态系统提供性能信息。 Anastasios强调了AI格局的一个重大转变：开源模型，特别是来自中国的模型，正在迅速改进。他提到中国模型Kimi K3最近在前端编码等特定任务中击败了顶级的美国闭源模型（包括Fable）。他表示，这一事件“违背了中国模型只是美国模型提炼品”的说法，并标志着模型商品化趋势。 他认为，美国迫切需要一家价值数千亿甚至万亿美元的公司，专注于“美国优先的开源”AI。这得益于企业寻求“AI主权”——即渴望拥有自己的整个AI供应链，用专有数据微调开源模型，并降低与外部供应商相关的风险。开源的新商业模式，例如与推理提供商进行收入分成，或提供“全栈开发和工程”（FDE）服务，正在成为可持续的途径。 关于众多的“新实验室”（估计超过75家），Anastasios持怀疑态度，预测其中三分之二最终将失败或被收购用于拆分。他警告说，虽然最初的估值可能基于团队的背景，但“下一轮融资将非常艰难”，需要营收的超高速增长，这是许多公司将难以实现的。 从地缘政治角度看，他承认中国在职业道德和政府支持方面的优势，但指出美国芯片出口管制是当前的障碍，尽管这可能无意中促使中国建立自己的生态系统。他探讨了出口管制的有效性，质疑“让世界依赖美国硬件”是否比饿死竞争对手是更好的长期策略。Anastasios预计，出于国家安全考虑和美国大型实验室的游说力量，美国未来可能会限制中国的开源模型，尽管这可能会损害美国企业的利益。他强调了AI模型中“后门”的危险性，即使是本地部署的模型，也可能通过特定提示被利用来泄露敏感数据，预计网络攻击将“疯狂”增加，包括AI生成的虚假求职者渗透公司。这需要新的招聘流程，例如当面入职以验证身份。 他认为，数据市场是AI模型的“规模化补充”，到2030年将增长到1000亿至1万亿美元。他驳斥了风险投资家对“营收集中度”的担忧是不必要的，并引用了像台积电这样在营收集中度高的情况下依然蓬勃发展的成功公司。他坚持认为，数据不像GPU那样是商品，而且对企业AI至关重要。Arena公司本身，年化营收超过1亿美元，正专注于智能体评估，他认为这是AI部署中的一个关键瓶颈。 Anastasios预见模型提供商将积极进入应用层，对现有的SaaS业务构成重大竞争风险。他还对AI在医学领域的潜力表示兴奋，特别是在根除慢性病方面，并强调了数据基础设施在实现这一目标中的关键作用。

Anastasios Anjapaloulos, founder and CEO of Arena, offered a blunt and insightful perspective on the rapidly evolving AI industry, particularly concerning model evaluation, geopolitical dynamics, and emerging business models. Arena, his company, is positioned as the central evaluation platform for AI, moving beyond static benchmarks to measure real-world AI performance. It assesses models based on factors like factual accuracy, steerability, human preference, and hallucination when put in the hands of real people. This direct feedback helps labs improve models and informs the ecosystem about performance. Anastasios highlighted a significant shift in the AI landscape: open-source models, especially from China, are rapidly improving. He cited Kimi K3, a Chinese model, which recently beat top American closed-source models (including Fable) in specific tasks like front-end coding. This event, he stated, "violates a narrative" that Chinese models are merely distilling American ones and signals a move towards model commoditization. He believes there's an impending need for a multi-hundred billion or trillion-dollar American company focused on "American first open source" AI. This is driven by enterprises seeking "AI sovereignty"—the desire to own their entire AI supply chain, fine-tune open-source models with proprietary data, and mitigate risks associated with external vendors. New business models for open source, such as revenue sharing with inference providers or offering "Full-stack Development and Engineering" (FDE) services, are emerging as sustainable paths. Regarding the numerous "Neolabs" (estimated at 75+), Anastasios is skeptical, predicting two-thirds will ultimately fail or be acquired for parts. While initial valuations may be based on the team's pedigree, he warns that "the next round's a bitch," requiring hyper-growth in revenue, which many will struggle to achieve. Geopolitically, he acknowledged China's advantages in work ethic and government support but pointed to US chip export controls as a current hindrance, though it could inadvertently incentivize China to build its own ecosystem. He debated the efficacy of export controls, questioning whether "addicting the world to American hardware" might be a better long-term strategy than starving competitors. Anastasios anticipates the US will likely restrict Chinese open-source models in the future, driven by national security concerns and the lobbying power of large American labs, despite the potential to disadvantage US businesses. He stressed the danger of "backdoors" in AI models, even locally hosted ones, which could be exploited through specific prompts to leak sensitive data, anticipating a "fucking insane" increase in cyberattacks, including AI-generated fake job applicants infiltrating companies. This necessitates new hiring processes, such as in-person onboarding, to verify identity. The data market, he argued, is a "scaling complement" to AI models, set to grow to $100 billion to $1 trillion by 2030. He dismissed venture capitalists' "revenue concentration" concerns as unwarranted, citing successful companies like TSMC that thrive despite it. Data, he maintained, is less of a commodity than GPUs and crucial for enterprise AI. Arena itself, with over $100 million in annualized revenue, is focusing on agentic evaluations, an area he views as a critical bottleneck in AI deployment. Anastasios foresees model providers aggressively moving into the application layer, posing a significant competitive risk to existing SaaS businesses. He also expressed excitement about AI's potential in medicine, particularly in eradicating chronic diseases, highlighting the critical role of data infrastructure in achieving this.