Jeff Dean joined Google in 1999 as the 30th employee, co-designing MapReduce, BigTable, and Spanner with Sanjay Ghemawat, laying the foundation for large-scale distributed systems, later leading TensorFlow and the early shift to neural networks, and has been core to the company's AI strategy for the past 15 years; Oriol Vinyals and Quoc Le collaborated with Ilya Sutskever in 2014 to propose seq2seq, establishing the modern encoder-decoder paradigm and participating in Gemini's technical leadership.

On the capital front, Radical Ventures and Khosla Ventures led the seed round, with Alphabet entering through equity and initial computing power, motivated to shift the four's accumulated expertise in systems and models from internal product iterations to independently scalable experimental automation, reallocating resources from Google's internal research budget to an external nonprofit company structure, allowing scientific closed loops to no longer be constrained by the pace of large model competitions.

Similar cases can be seen in DeepMind's transition from an independent lab to being acquired by Google, and OpenAI's early shift from nonprofit to commercialization; currently, Discovery Loop is in a transformation phase from infrastructure contributor to scientific discovery control layer, and the departure of the four further exacerbates the recent outflow of AI talent from Google.

Essentially, this represents a technological replacement: substituting traditional human scientist-dependent sequential experimental cycles with parallel, scalable AI-driven closed loops, where the mechanism lies in the combination of high-citation researchers' system expertise with capital, enabling higher throughput iterations that break through the limitations of a single company's internal resources and priorities.

ABAB News · Cognitive Laws

Infrastructure leaves, the next round of discoveries begins
Talent outflow = platform deregulation × new closed loop initiation
Scientific acceleration relies not on larger models, but on shorter experimental cycles