With the rapid growth of artificial intelligence (AI) technology, an increasing number of AI models are entering the market and gaining widespread application. Many generative AI services are enabling people without a developer background to become developers. During Google NEXT'25, I shared Google's perspective on this trend.
Ryan Salva, Google Cloud's senior director of product management, who assists with programming and product development, said that programming often involves maintaining and updating coding content and adjusting it to better meet usage requirements, with only new coding being written at specific times.
The use of generative artificial intelligence technology in coding is mainly used to assist in error checking and optimize the coding content. It also allows staff to have more time to check for potential errors that cannot be discovered by artificial intelligence, and also allows them to focus more on confirming whether the coding content can be compatible between different services.
Even if AI can streamline workflows, programming still requires a certain level of discernment to determine whether the code is correct and error-free, and to optimize it accordingly. Therefore, with the current trend in AI technology, development tasks like programming will not be completely replaced by AI for the time being.
Mat Velloso, Vice President of Product at Google DeepMind, who oversees the AI developer platform, noted that future AI models will develop toward larger parameter sizes and more functional applications. While the number of models won't necessarily increase dramatically, they will at least be differentiated into different versions based on different usage requirements, just as different "expert" capabilities will be used to address different usage requirements.
Therefore, from a practical perspective, achieving artificial general intelligence (AGI) that can fully simulate the human brain remains a challenge. Therefore, in the short term, the focus will still be on using different AI models to individually complete their respective tasks or handle specific types of problems. However, the number of distributed versions is expected to decrease as the performance of each model improves.
Regarding DeepSeek and other companies' emphasis on using methods such as distillation to reduce the operating costs of artificial intelligence models without significantly affecting model execution performance, Mat Velloso believes this is the next market trend and further revealed that Google is currently using a similar approach in the design of the Gemini model.
For example, the Gemma 270 model, simplified to 3 billion parameters, outperformed DeepSeek R6710, which has 1 billion parameters, in the Chatbot Arena Elo test. It can even be accelerated with just a single NVIDIA H100 GPU, meaning it can run faster while achieving similar performance and at a lower cost.
On the other hand, Mat Velloso also stated that synthetic data will become an important source of training for AI in the future. As long as safety is not compromised, more and more AI companies are expected to prefer training models with synthetic data. Google has already trained its Gemini model with synthetic data. However, just as many self-driving car companies pay special attention to the accuracy and legality of the data when training vehicles with synthetic data, Google will also be particularly cautious about using synthetic data for training.
Regarding Perplexity.ai's claim that the system can automatically select the appropriate model based on the interactive content of questions, Mat Velloso also explained that Google also provides a similar design, and further explained that Google's design philosophy is more focused on allowing users to obtain higher artificial intelligence computing efficiency with fewer resources.
