AI AGENTS: 1000X Productivity! Companies MUST Invest Now! #shorts

这段文字强调了AI代理的变革性力量，使个人能够实现10倍、100倍甚至1000倍的生产力提升。 核心论点是，只要生产力回报超过了成本，拥有充足资源的公司就应该大力鼓励其工程师在这些AI工具上投入“一切必要资源”。这种投资应与工程师的薪资挂钩。 英伟达首席执行官黄仁勋被引用（虽半开玩笑，但意图严肃地）说，如果他每年支付“50万美元”薪水的工程师没有至少在AI代币（例如，用于编码代理）上花费等量的资金，他会“非常非常生气”，并质疑他们为何不利用AI代理来实现极致的生产力。 演讲者认为，任何向工程师支付“每年几十万美元”薪资的企业，如果不为他们提供必要的“代币”以“将其产出翻倍、三倍、五倍甚至十倍”，那将是“愚蠢的”，因为这会带来“巨大的多倍”投资回报。 这一战略对于科技公司尤为重要，在这些公司中，“工程师的能力以及他们完成工作的速度”往往是一个“限制因素”。这些公司通常“资金充裕”、“利润丰厚”、“财源滚滚”，这使它们处于有利地位，可以投资于这种“巨大”的生产力潜力。 一个说明性的例子是已知的“首个AI计算交易”，涉及**Cursor和SpaceX**。Cursor暂时付费使用了SpaceX“Colossus”计算资源的一部分来训练其模型。这催生了“Composer 2.5”，它在“帕累托尺度”上实现了“每成本约一个数量级的性能提升”。Composer 2.5“几乎与地球上最好的编码代理一样强大”（相似度在5%或10%以内），但使用成本却“大约是十分之一”，却能实现“几乎相同的性能”。 这种成本效益促使用户重新评估，促使他们考虑，如果这意味着在相同预算下能够使用“10倍的代币”，那么稍逊于“神级”的能力是否更可取。 总结来说，Cursor在SpaceX的计算资源上进行了“最多几个月的训练”，从而开发出了一款“性价比提高10倍”的新模型，并获得了广泛采用。这一成功最终导致**SpaceX收购了Cursor**，实际上，由于最初在计算资源上的投资，使得Cursor的产品“变得好10倍”。 这段文字最后强调了，从为*训练*AI模型（而不仅仅是用于推理，即执行任务）获取计算资源所获得的投资回报是“荒谬的”。

This transcription emphasizes the transformative power of AI agents, enabling individuals to become "10 times, 100 times, or even 1,000 times more productive." The core argument is that companies with sufficient resources should aggressively encourage their engineers to invest "whatever is necessary" in these AI tools, provided the return in productivity outweighs the cost. This investment should be viewed in relation to an engineer's salary. Jensen Huang, CEO of NVIDIA, was cited (semi-jokingly, but with serious intent) stating that he would be "real fucking mad" if an engineer he pays "half a million dollars a year" was not spending at least that much on AI tokens (e.g., for a coding agent), questioning why they aren't leveraging agents for extreme productivity. The speaker argues that any business paying engineers "a few hundred thousand dollars a year" would be "moronic" not to provide them with the necessary "tokens" to "double or triple or 5x or 10x their output," as this yields a "massive multiple" return on investment. This strategy is particularly relevant for tech companies, where the "capability and how quickly their engineers can get stuff done" is often a "limiting factor." These companies are typically "flush with cash," "highly profitable," and "print money," making them well-positioned to invest in this "enormous" potential for productivity. An illustrative example is the "first AI compute deal" known, involving **Cursor and SpaceX**. Cursor temporarily paid for access to a portion of SpaceX's "Colossus" compute to train their model. This resulted in "Composer 2.5," which achieved "roughly an order of magnitude increase in capability per cost" (on the "Pareto scale"). Composer 2.5 was "almost as capable as the best coding agents on Earth" (within 5% or 10% similarity) but cost "roughly a tenth the cost" to use for "almost identical performance." This cost-efficiency led to a re-evaluation by users, prompting them to consider if slightly less than "godlike" capabilities were preferable if it meant being able to use "10 times as many tokens" for the same budget. In summary, Cursor used a "couple of months at most of training" on SpaceX's compute, resulting in a new model that offered "10 times more bang for buck," leading to widespread adoption. This success culminated in **SpaceX acquiring Cursor**, effectively making Cursor's product "10 times better" due to the initial investment in compute. The transcript concludes by highlighting the "absurd" return on investment derived from access to compute for *training* AI models, not just for inference (doing the thing).