Summary of Key Points
Four leading figures in Google's AI field (Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals) have collectively left to founding a new company called Discovery Loop. Among them, Oriol Vinyals and the previously departed Noam Shazeer, along with Jeff Dean (the head of Gemini's technology), were the three co-technological leaders of Gemini. All three left within seven weeks. Google's stock price fell by more than 5% on the same day. Koray Kavukcuoglu was appointed to take over Gemini's research and development efforts. This change will have little immediate impact on Gemini 4, but in the long run, it may weaken Google's ability to develop original architectures, its full-stack collaboration capabilities, and its organizational appeal. It also exposes the internal contradiction at Google: the company wants to release products quickly while maintaining long-term innovation, a balance that is difficult to achieve.
I. The Departures Are Not Individual Transfers, but a Collective Exodus of the Core Team
The key issue here is not just the departure of Jeff Dean; rather, it represents the near-complete depletion of Gemini's technical core:
- Within seven weeks, all three co-technological leaders of Gemini left: Noam Shazeer joined OpenAI in June, and Jeff Dean and Oriol Vinyals started their own business in August.
- The four founders of the new company are all fundamental figures at Google: Jeff Dean (who laid the foundation for Google's AI hardware and software, such as TPU and TensorFlow), Sanjay Ghemawat (who helped build Google's cloud computing infrastructure), Quoc Le (one of the co-inventors of Transformer), and Oriol Vinyals (the direct technical leader of Gemini and the creator of AlphaStar).
This is not a routine personnel change; it's the departure of the team that understands the entire AI ecosystem from chips to models, essentially taking away half of Google's "AI brain design."
II. Short-term Impact: Gemini 4 Will Not Be Delayed, but Existing Issues Were Already Present
Many are concerned that Gemini 4 might be delayed, but the impact is limited:
- Gemini 4 is already in the pre-training phase: The training methods and data for the model have been determined. The engineers and researchers responsible for implementation remain, and Koray Kavukcuoglu, the new leader, is a former DeepMind employee who has worked on core projects like WaveNet and DQN, so he is not an outsider.
- Current issues are unrelated to the departures: The release of Gemini 3.5 Pro was postponed twice (originally scheduled for June but still not out), due to internal quality issues—this problem existed before the departures, indicating that Google has long had difficulties with product delivery.
In short, Gemini 4 will be released on time, but the delays with 3.5 Pro highlight Google's ongoing issue of being unable to accelerate its progress.
III. Long-term Implications: Google May Lose Its Ability to Innovate New Tools
While there are no immediate problems, the long-term consequences are significant:
1. Loss of Architectural Expertise: Figures like Shazeer and Quoc Le are not just skilled at fine-tuning models; they invented fundamental AI structures such as Transformer, sparse MoE, and seq2seq, which are now widely used in the industry. This kind of expertise is based on years of experience and intuition and cannot be easily documented or transferred. It will be difficult for Google to find replacements.
2. Weakening of Full-stack Collaboration: Google's strength lies in its ability to develop everything from chips (TPUs) to models (Gemini) in-house. With Dean and Ghemawat gone, this capability is at risk.
3. Deterioration of Organizational Cohesion: Vinyals was recruited by Dean; now that he has left, senior researchers may question the value of staying at Google. Google has already lost many AI talents this year, and this departure could accelerate this trend.
IV. Google's Response: "Hedging Risks," but Admitting Internal Limitations?
Google's approach is strategic yet somewhat helpless:
- It has become a founding investor in Discovery Loop and provides computing power through its cloud services. This serves as a hedge: if the new company succeeds, Google can benefit; if it fails, the loss is minimal.
- However, this implies that Google internally lacks the ability to rapidly advance autonomous AI research. The fact that these four individuals, who understand Google's limitations best, chose to leave suggests that there is insufficient space within the company for such long-term, non-profit-oriented initiatives.
Dean stated, "Outside a public company, we can make decisions that do not solely align with financial interests." In other words, within Google, short-term profitability must take precedence over long-term innovation.
V. The Future of Gemini
The success of Gemini depends on several factors:
- Short-term (Gemini 4): There should be no issues with its release, but the quality remains to be seen. Will it meet expectations or just be a follow-up to previous versions?
- Medium-term (Gemini 5/6): Will new architectures be developed? If Google can only improve existing models by tweaking parameters and accumulating more data, it will fall behind companies like OpenAI and Anthropic.
- Long-term: Will Google retain its full-stack advantage? Without the ability to develop from chips to models in-house, Gemini's strength could be compromised.
These departures are not the cause of Gemini's failure, but they have brought to light existing issues. Google holds significant advantages (TPUs, cloud, users), yet it often lags behind. With its most innovative talents gone, the question arises: can Google still play this game effectively? Gemini 4 will provide an initial answer, but the people responsible for its success are no longer those who designed the project.
In summary, this event does not mean the end of Google's AI efforts; rather, it may mark the end of an era of genius-driven innovation. Google now faces the challenge of balancing rapid product delivery with long-term innovation, a difficult balance to maintain. After all, breakthroughs like Transformer and TPU were not the result of strict schedules or plans.