Summary of Key Issues
Google's AI efforts have recently encountered twin challenges: Firstly, the development of its Gemini series of models has fallen short of expectations, with multiple delays and performances that cannot match those of competitors like OpenAI and Anthropic. Secondly, there has been significant turmoil within the core AI department, DeepMind, involving the departure of top scientists and a reorganization of management. The root of these problems lies in Google's "full-stack strategy," which has created organizational conflicts. As a once-autonomous AGI (Artificial General Intelligence) research institution, DeepMind now has to align with Google's overall product roadmap and resource allocation, leading to a loss of research autonomy and increased internal competition for resources, ultimately affecting the speed of model development and innovation capabilities.
1. Model Development Lag: Gemini Series Falls Behind Competitors
Google's Gemini models were once highly anticipated but have continuously failed to meet expectations:
- Repeated Delays in Release: The development of Gemini 4 has been slow, with two postponements; Gemini 3.5 Pro was also scheduled for release in June and July but missed both deadlines.
- Inadequate Performance: The released versions, Gemini 3.5 Flash and 3.6 Flash, do not surpass the performance of OpenAI's GPT-4 or Anthropic's Claude 3, and even lag behind domestic models like Kimi and Tongyi Qianwen.
- Changing Public Opinion: While Google's Gemini 3 once caused OpenAI to establish a "red team" in response, it is now viewed as further behind in innovation.
In short, Google's flagship AI products are not only late to market but also inferior to those of its competitors.
2. Personnel Turmoil: Top Talents Leaving, DeepMind's Authority Restrained
The recent changes at DeepMind are quite significant:
- Departure of Key Scientists: Former Chief Scientist Jeff Dean has left to start a new company called Discovery Loop, which aims to use AI to solve scientific and engineering problems (problems that DeepMind could have tackled before but no longer can).
- Management Transition: DeepMind's founder, Demis Hassabis, stepped down as CEO to become Chairman and Chief Scientist at Alphabet, retaining his scientific authority but losing management control. The former CTO, Koray, was promoted to Senior Vice President, responsible for daily operations and reporting directly to Google CEO Sundar Pichai, making DeepMind more aligned with the parent company's directives.
These changes are not incidental; DeepMind's previous autonomy has been diminished, prompting scientists to seek new opportunities. Google, on the other hand, hopes to align DeepMind more closely with its product priorities through management adjustments.
3. The Double-Edged Sword of the Full-Stack Strategy
Google's full-stack approach, which covers everything from chips and cloud services to models and applications, was once a strength (enabling customers to purchase comprehensive solutions), but it has now become a source of internal conflicts:
- Fierce Competition for Resources: There is intense competition for computing resources between Google's cloud division (which sells computing power to generate revenue), DeepMind (which uses it for model training), and the search division (which optimizes products). The cloud division prefers to sell its resources to large clients like Anthropic, which offers stable income, while DeepMind's investment in model training is less profitable.
- Unequal Commercial Channels: Google Cloud controls the commercial sales of Gemini models, preventing DeepMind from selling them directly. Conversely, Google Cloud can sell computing power to Gemini's competitors, effectively working against DeepMind.
- Distributed Product Ownership: The development of Gemini's foundational models, as well as individual and enterprise applications, is spread across DeepMind, Google Labs, and Google Cloud, leading to delayed feedback. For example, improving code capabilities requires data from corporate clients, which are held by Google Cloud, giving DeepMind a disadvantage compared to competitors like Anthropic, which have integrated product and model development.
The full-stack strategy, while beneficial in some respects, has created internal divisions over resources.
4. DeepMind's Loss of "Soul": From AGI Laboratory to Product Tool
DeepMind was originally a unique entity acquired by Google as an autonomous research institution focused on AGI, allowing scientists to pursue high-risk, long-term research. However, this has changed:
- Restricted Research Directions: Google now requires DeepMind to focus on areas where it has weaknesses (such as code capabilities) and on Google's core products (search and cloud services), limiting its ability to explore unproven, cutting-edge directions.
- Pressure for Rapid Updates: CEO Pichai demands monthly model releases, which maintains a pace but stifles breakthrough research. Major innovations often require time and experimentation, unlike the success of AlphaGo, which was achieved through continuous iteration.
DeepMind has transformed from a creative scientific team to a department driven by KPIs, potentially leading to the erosion of its innovative capabilities.
5. Google's Dilemma: Short-Term Product Momentum vs Long-Term Innovation
Google faces a choice:
- Short Term: By focusing on immediate product releases, Google ensures its presence in the AI market and generates revenue from its cloud business. However, this sacrifices long-term breakthroughs and may prevent it from catching up with competitors like OpenAI's GPT-5 or Anthropic's next-generation models.
- Long Term: Allowing DeepMind to operate autonomously could lead to slower model releases and resource inefficiencies, impacting current commercial interests.
Pichai has chosen the former approach, believing that a more controlled and focused DeepMind is more valuable. However, this may result in Google falling behind in the future AI race.
In summary, Google's AI challenges reflect the inherent issues of large companies facing rapid AI innovation: internal conflicts stemming from its full-stack strategy, the transition of DeepMind from a research institution to a product-oriented entity, and the tension between short-term commercial goals and long-term innovation. If these issues are not resolved, Google could lose its leading position in the AI landscape.