AI's REAL Opportunity: Almost-Einstein Models Explained #shorts

演讲者认为，AI模型的主要机会并不在于实现智能的绝对顶峰或前沿模型，这些模型被他们称作“绝对的爱因斯坦”。尽管承认在“前沿模型的最前沿”存在“非常具体的用例”需要这种“顶尖智能”，但这些用例被认为是“相当有限的”。对于大多数任务而言，用户并不需要“世界上最顶尖的智能”。 相反，“接近爱因斯坦”水平的性能被认为“对大多数用例来说绰绰有余”，因为大多数模型对于大多数任务“不需要是字面意义上的爱因斯坦”。尤其如此，当这种“接近爱因斯坦水平的性能”的成本，相当于“支付麦当劳的工资”——只是绝对最佳模型“成本的一小部分”时。 因此，真正的机会在于那些“几乎和最好的模型一样好”的模型，它们的性能“与最顶尖模型的能力相差无几，仅在几个百分点之内，或许是百分之五以内”，但“成本却只是冰山一角”。这些模型被描述为“工蜂模型”或“苦力模型”。 假设这些模型能够保持“与绝对最佳前沿模型几乎一样好”的能力，但能以显著降低的成本完成工作，演讲者总结说，“大部分AI收入似乎不可避免地会花在那些超高效的‘接近爱因斯坦’水平的模型上。”

The speaker argues that the major opportunity with AI models does not lie in achieving the absolute pinnacle of intelligence or frontier models, which they refer to as "the absolute Einsteins." While acknowledging there are "very specific use cases" for such "pinnacle intelligence" at the "frontier of frontier models," these are deemed "fairly limited." For the majority of tasks, users do not require "the biggest brain on the fucking planet." Instead, "almost Einstein" level performance is considered "more than enough for most use cases," as most models for most tasks "don't need to be literal einstein." This is particularly true when such "almost Einstein level performance" comes at the equivalent of "paying a McDonald's salary" – a "fraction the cost" of the absolute best. The real opportunity, therefore, is in models that are "almost as good as the best," performing "within just a few percent, maybe within five percent of the performance capability of the absolute best of the best models," but at a "fraction the cost." These are characterized as "worker bee models" or "the grunts." Assuming these models can maintain capability "almost as good as the absolute best frontier models" but are able to do their work for a significantly reduced cost, the speaker concludes that "most of the AI revenue seems inevitably likely to be spent on the hyper efficient almost einstein level models."