Time is a cruel game, such as when the stunning Apple Vision Pro finally appeared, no one remembered the distant Nokia.
But all mobile games, including Ryan McLeod, the game developer who will come to Apple Vision Pro in the future, need to pay tribute to Nokia's decision to build Snake in the first place in 1997, a move that really opened up mobile gaming history.
The game logic of "Snake" is very simple, eating fruit can score points, but the body will become longer, and the direction of control will not hit the boundary or yourself. Today, Nokia has almost left the public eye, the game with its roots in Snake is still active, and even "how to write a snake game in the shortest code" is still an attractive topic in the developer community.
This now classic and aesthetically pleasing game has also become a ruler for measuring AI capabilities in the era of large models. Shi En, deputy general manager of Baidu Intelligent Cloud AI Platform, with the help of a new code assistant Comate, started to build "Snake" from the canvas, until finally Snake twisted and appeared, and AI completed most of the code work.
The whole process takes less than 5 minutes.
Comate for developers
The short but intense competition of large models Until now, the fog of the black box has not lifted, and the attention of the outside world is changing. The grand visions begin to calm down and translate into a focus on more concrete things. For example, what new productivity can generative AI bring to the industry today?
On June 6, at the Wenxin Big Model Technology Exchange Conference held in Chengdu, Baidu opened the code assistant Comate for testing. This is an intelligent development tool similar to code assistants such as GitHub Copilot, but with more Chinese annotations and development documentation as training data. During the coding process, Comate can deduce possible next input choices based on what the developer is currently writing.
The Wen Xin big model is grand, and Comate is concrete.
For the game development (web version) process of "Snake", the developer must first draw the bottom on the canvas, and then set the operation mode of the keyboard and the judgment of the game end conditions, in addition, the control of environmental elements such as crawling speed. For Comate, you only need to enter "canvas", and Chinese notes of "flexible layout, horizontal center, vertical center" to complete the generation of the game canvas code. Then enter simple words such as "param color", "left", "food=" and other colors, directions, and food, and Comate will automatically understand the instructions in relation to the context, complete the code, and switch between multiple recommended codes, and after selecting the appropriate code, the runnable "Snake" game is directly generated.
According to Baidu, Comate's capabilities have been integrated into all Baidu's business lines for the first time and have achieved good results: 50% of the code in the core R&D department can be generated by Comate. Comate has been conducting extensive testing inside Baidu since last year. The test results show that nearly 50% of the code written with the assistance of "Comate" has been adopted by developers, and it has been widely used in various product development within Baidu.
In addition to Baidu's internal codebase, the objects of Comate Deep Learning also include high-quality Github codebase. Now this code assistant can realize a variety of intelligent functions such as automatic code generation, code autocompletion, code intelligent search, high-quality code recommendation, and automatic test code generation. The inference speed can achieve about 300ms for a single request, which means that developers do not need to stop and wait for code generation, and Comate can fully match the coding speed of developers.
At present, Comate code assistant supports more than 30 mainstream development languages/frameworks, and has specially optimized data for C/C++, Java, Python and other mainstream languages Comate to achieve better code recommendation effects. In addition, Comate also supports front-end and back-end, software and hardware scenarios, and a variety of IDEs commonly used by programmers.
From this point of view, "Snake" is an intuitive but inadequate demonstration. Shi En also said that the code development of "Snake" can already rely entirely on large models to automatically generate without interference. It's just interesting that Comate has benefited from the ability of large models from a "code recommendation tool" launched by Baidu's internal intelligent work platform to the current "code generation assistant".
Shi En, deputy general manager of Baidu Intelligent Cloud AI Platform Source: Baidu
The incubation of Comate began around 2018, and Baidu has already mentioned code generation when looking for ways to improve development efficiency. However, the technology is not mature enough, so first choose to improve the efficiency by retrieving the algorithm and making algorithm recommendations. "At that time, technology exploration was already being attempted. But after the emergence of the Wenxin model, we applied the real code generation in a wider range of scenarios," Li Jingqiu, deputy general manager of Baidu Intelligent Cloud AI Platform, told Pinplay.
Comate transitioned from search logic to generation logic, and the ability of large models began to show a productivity transformation posture in the development process. And if Comate is an answer of the Wenxin model to developers, then "Wen Xin Yiyan - Turbo" is a new solution handed over by Baidu to the industry two months after Wenxin Yiyan was launched. Hidden behind is the iteration route of the Wen Xin Qianfan model platform gradually clear after two months.
The iterative direction of Wen Xin Qianfan: effect + efficiency
Baidu defines Wenxin Qianfan as the world's first one-stop enterprise-level large model platform. Specifically, Wenxin Qianfan not only provides large model services including Wen Xin Yiyan's underlying model (Ernie bot), but also provides a variety of AI development toolchains and a complete set of development environments. In addition, the big model platform also supports various third-party open source and closed source large models. Since the first launch of the internal beta on March 27, with the support of Wen Xin Qianfan, Wen Xin Yiyan has completed four technical iterations in two months.
At a technical exchange meeting on April 25, Hou Zhenyu, vice president of Baidu Group, revealed that since the internal test, through the continuous optimization of algorithms and models, Wen Xin's reasoning efficiency has increased by 10 times, and the reasoning cost has been reduced to one-tenth of the original. A month and a half later, Wen Xin Yiyan's high-performance mode "Wen Xin Yiyan - Turbo" appeared as the epilogue of the first iteration of Wen Xin Qianfan. In some high-frequency and core scenarios, the overall performance of the inference service has been improved by a total of 50 times while meeting the same customer needs.
After many iterations, Wen Xin Qianfan's two evolutionary directions have also emerged: effect and efficiency.
In terms of effect, in addition to the significant improvement of inference performance, "Wen Xin Yiyan - Turbo" supports SFT training, and provides a variety of training methods for different scenarios and effects, and Bloom7B (7 billion parameters) third-party large model can support different training methods such as P-tuning, SFT, and Lora. And due to the needs of enterprises for large model retraining and the consideration of privatization deployment, Wenxin Qianfan will open the plug-in protocol, which can allow third-party enterprises to share plug-ins based on the plug-in protocol, and fully access the internal data through the plug-in method, while achieving better large model effects.
In addition, Wenxin Qianfan will provide some prefabricated Prompt templates, as well as support for adding, deleting, modifying and checking Prompt templates, and calling Prompt templates through service interfaces to obtain better reasoning effects.
Parallel to the effect is efficiency, or it can also be understood as cost performance.
The 50-fold increase in the inference efficiency of "Wen Xin Yiyan - Turbo" means a significant decrease in the cost of large model reasoning, which will be another improvement after the internal testing of Wen Xin Yiyan large model after only one month and the cost of large model reasoning dropped to 1/10 of the original. In terms of data labeling, "Wen Xin Yiyan - Turbo" can generate annotation data in batches in the future and quickly use it for subsequent training, which will greatly reduce the cost of data labeling compared with the previous manual data annotation form.
In terms of deployment methods, Wen Xin Qianfan is also gradually increasing its flexibility to respond to customers.
Wenxin Qianfan provides "3+3" delivery methods. On the public cloud service side, three services are provided (directly calling the inference capabilities of large models), fine-tuning (efficiently training large models in specific industries through high-quality and fine-standard business data), and hosting (publishing models on Baidu Intelligent Cloud to achieve more stable and efficient operation) to reduce the threshold for enterprises to deploy large models.
On the privatization deployment side, which has attracted much attention, Wenxin Qianfan supports software licensing (providing large model services running in an enterprise environment), software and hardware integration (providing a complete set of large model services and corresponding hardware infrastructure) and leasing services (providing the leasing of machines and platforms to meet the low-frequency needs of customers). The latest addition of rental services is designed to cater to customers on a budget or with less frequent model training needs.
The rapid iteration of the Wenxin Qianfan model platform in two months, as well as the clear route planning at both ends of the effect and efficiency, are not only the advantages of computing power, but also reflect the overall advantages of Baidu's four-layer AI architecture.
Li Yanhong Source: Sohu
Robin Li has publicly stated that Baidu is the only artificial intelligence company in the world that carries out a full-stack layout in the four layers of chips, frameworks, models and applications. As Baidu's layout on the underlying chip, Kunlun Core has achieved the deployment of tens of thousands of pieces of two generations of products, and the third generation is expected to be mass-produced early next year; Flypaddle is already the deep learning framework with the first comprehensive market share in China; In addition to Wenxin, there are also general large models including NLP, CV, cross-modal and 11 industry large models, and the rich accumulation of application layers is a data flywheel. This complete layout also gives Qianfan the prerequisites for flexible evolution in the face of 300 ecological partners and more than 400 internal scenarios.
Every technology evolution is the joint efforts of multiple teams from the chip layer to the framework layer, to the model layer and the application layer. "We will pull a working group internally, there are people from Kunlun chips, people from the flying paddle framework team, people with large-model NLP algorithm strategies, people from our large-model platform tool chain, and people from terminal application intelligent customer service." Each month, we set common goals, such as how far performance must be improved this month. Li Jingqiu said.
As Hou Zhenyu, vice president of Baidu Group, said at a closed-door meeting in early May, "In the era of big models, the innovation of enterprises needs not only intelligent computing power, flexible framework platforms, rich large model systems and high-quality application solutions, but also end-to-end adaptation and optimization between these four, which is an end-to-end innovation project that 'wants, wants, and wants'." ”
And almost everyone can feel that Baidu is beginning to show its stamina in the battle of large models.