Jeff Dean & Noam Shazeer – 25 years at Google: from PageRank to AGI (AI column)
Interesting. interview Two legendary Google engineers uncover the company’s transformation from a startup to a leader in artificial intelligence. Jeff Dean and Noam Shazier, who worked in the corporation together 50 Over the years, they shared insights on technological evolution, current projects and the future of AI.
The evolution of Google and personal experience Company growth 25 Employees before the tech giant was accompanied by a paradigm shift in management. Dean recalls the stages, from personal acquaintance with all colleagues to management system through project metrics. Shazir highlights the uniqueness of Google's early culture, where he began under the mentorship of Dean himself, who developed the company's underlying infrastructure. Noam left the company twice. (roughly 12 years)Which became an internal joke. His last comeback cost the corporation $2.7 Billion under the terms of the deal with Character.AI.
Development of AI technologies The main attention in the interview is paid to the project Gemini – Google’s flagship AI system. The developers reveal three key aspects:
- Architectural innovation: transition from monolithic models to a modular system with asynchronous updating of components. This allows you to improve individual modules without completely retraining the system.
- Context window: work on expanding to trillions of tokens, which is many times higher than current figures. Experiments show that models with increased context demonstrate elements of long-term planning.
- The economy of computingThe cost of AI operations is reduced in 100 Compared to reading an e-book and a million times to hiring a developer. This is due to the new TPUs. 5It is a generation that is optimized for inference.
Practical application 25A percentage of Google’s code is already generated by AI through a special version of Gemini, trained on an internal code base. MENA is a corporate chatbot, the predecessor of ChatGPT, used by employees with 2021 years. Shazir notes the paradox: despite public restraint, Google has been investing in AI for decades, starting with pioneering work on statistical machine translation and web-index-based spall-checking. 2000-x.
Future directions
- "Organic" architecturedevelopment of self-adapted systems where different modules specialize in specific tasks (language, code, planning).
- Multi-agent systemsManaging thousands of parallel agents with personalization through personal data integration (email, documents, health metrics). Hardware innovationChip design where AI automates 80% of the development process, reducing the cycle 3 years 9 months.
Engineers are sceptical about the “magic of big data”, emphasizing a systems approach: the success of modern models is the result of the synergy of algorithms, architecture and hardware optimization. Dean sums up the interview noting that we are at the beginning of a new era where AI will be an extension of human intelligence, not a replacement for it.
#AI #Engineering #Software #History #Management #Architecture