LLM large-scale models and generative AI will explode rapidly in the first half of 2023. Not only Goldman Sachs and McKinsey have released economic forecasts for generative AI, they all believe that generative AI will significantly improve global productivity and bring global With the economic growth of trillions of dollars, UC Berkeley and Stanford have successively released the LLM large-scale model ranking list. The world's top LLM large-scale models have reached nearly 30, and this does not include the many LLM large-scale models that have emerged in the Chinese market. .
With the approach of the LLM "100-model war", industry users need to develop customized large models and generative AI applications for the industry and enterprises based on the existing LLM large models more quickly. At the 2023 World Artificial Intelligence Conference (WAIC) on July 6, 2023, Appen, which has cooperated with the world's leading AI companies for more than 27 years, launched an intelligent LLM large model development platform and announced a strategic upgrade for generative AI—— "Break the circle" from AI data services to full-stack AI services.
Appen Intelligent LLM large model development platform provides industry-oriented AI with a set of large model data preparation, model training, model reasoning, and model deployment applications, covering all aspects of data set management, data labeling, computing resource scheduling, model evaluation, and model fine-tuning. Stack capabilities help enterprises easily embrace large models, build generative AI applications, and achieve transformative experiences for end users. In addition, Appen LLM product line also includes basic data, baseline model, model evaluation & fine-tuning, application development and other full-chain products, platforms and services.
(Dr. Tian Xiaopeng, Global Senior Vice President of Appen, General Manager of Greater China and North Asia)
"In the second half of 2023, the LLM large-scale model and generative AI market will show a huge burst and growth trend. The era of generative AI has just begun," Appen Global Senior Vice President, General Manager of Greater China and North Asia The manager Dr. Tian Xiaopeng said. "Appen focuses on high-efficiency economic mass production industry large models and generative AI applications, fully empowering the intelligent transformation of various industries!"
Time for a strategic upgrade
The first half of 2023 can be said to be the "Spring and Autumn and Warring States" period of LLM basic large models. Many basic large models have been born from all over the world to China, and more teams are running and entering the research and development of basic large models. As of the end of June 2023, 85 large-scale models from China have been collected on Github—mainly LLM basic large-scale models, as well as some large-scale models for industries and specific fields, plus the world's top basic large-scale models, " "Hundred Models Battle" is no longer an exaggerated expression.
The "Hundred Models War" is still in fierce battle, but it has also successfully allowed AI to break through the technology circle, which has aroused the attention and attention of a wider range of people. Globally, according to the annual CEO survey released by the IBM Institute for Business Value at the end of June, three-quarters of the interviewed CEOs believe that companies with the most advanced generative artificial intelligence will have a competitive advantage. In China, Gartner China Enterprise Artificial Intelligence Trend Wave 3.0 pointed out that Chinese enterprises are shifting artificial intelligence projects from prototypes to production. Most enterprises are no longer obsessed with why they need AI capabilities, but pay more attention to the construction of AI engineering capabilities.
In the second half of 2023, the trend of industrialized mass production and engineering of LLM large-scale models is clearly emerging from the "100-model war". In particular, many industries and corporate customers pay more attention to how to choose existing LLM large-scale models And fine-tuned to adapt to the business scenarios of the industry and enterprises, and truly use AI to improve the productivity of industries and enterprises. Simply understood, it is to efficiently and economically mass-produce industry LLM large-scale models and implement end-to-end industry large-scale models in enterprises to truly improve productivity. This will be the focus of the AI market in the second half of 2023.
At the beginning of 2023, Appen, which has been committed to providing high-quality labeled data for AI companies and enterprise AI for a long time, assessed the situation and boldly "bet", and comprehensively launched the company's strategic upgrade-from AI data services to full-stack AI services. To become an AI service provider for vertical industries. Armughan Ahmad, Appen's new global CEO, stated in the company's 2022 annual report that AI data labeling is the foundation, and the full-stack AI service represented by generative AI is the growth S-curve, and it is also Appen's next strategic focus. Appen already has powerful AI data labeling tools, platforms, and services. The next step is to quickly promote industrialized mass production industry models and generative AI applications on this basis, and open up trillions of new economic volumes.
Large model development in one stop
If you want to quickly industrialize and mass-produce LLM large-scale models and generative AI applications in the industry, you need an industry-oriented large-scale model development platform. This is also the hot spot and focus of the AI market in the second half of 2023. In the first half of 2023, some technology companies have launched industry-oriented large-scale model customization development or solutions, while specialized third-party large-scale model development platforms and end-to-end AI large-scale model development services are still blank in the market.
Appen Smart LLM Large Model Development Platform is developed by the Appen China team. It is a development platform for LLM large model fine-tuning (Fine-tune). Fine-tune large models of industries and business scenarios such as games, medical care, and customer service. Zhou Bo, product manager of Appen China, introduced that the Appen Smart LLM large-scale model development platform includes three modules: data, model and computing resource management.
(Architecture Diagram of Appen Intelligent LLM Large Model Development Platform)
For the development of LLM large models, high-quality labeled data is critical. The reason why ChatGPT can stand out is the introduction of high-quality manual annotation data. The analysis of UC Berkeley's LLM leaderboard found that high-quality fine-tuning data sets are more important than model size, especially managing high-quality data sets in the pre-training and fine-tuning stages is a key method to reduce the size of the model while maintaining the high quality of the model . More and more studies have found that high-quality labeled data is very important or even a key method for model fine-tuning results and reducing the size of the model while maintaining the quality of the model.
The data module of Appen Intelligent LLM large model development platform comes from another popular product of Appen China: MatrixGo enterprise-level high-precision data labeling platform. MatrixGo is an enterprise-level platform for deep learning and machine learning data labeling. It not only has powerful Labeling tool set, AI-assisted labeling, flexible and visual workflow, and integration of Open API and external data platforms and closed-loop data. In response to the development needs of LLM, the Appen China development team combined MatrixGo technology to develop the data module of the LLM large model development platform, which can ensure the quality and efficiency of data labeling, while continuously reducing the cost of labeling.
The data module of Appen Intelligent LLM large-scale model development platform includes dataset management and data collection annotation, among which: dataset management includes data processing, data retrieval, data visualization, data slicing and other functions; data collection annotation includes personnel management, workflow engine , labeling tool engine and automatic labeling algorithm and other functions.
The core of the Appen Smart LLM large model development platform is the model module, which includes three parts: model evaluation, model fine-tuning and model deployment, among which: model evaluation provides A/B testing, standard corpus testing, custom testing, test result visualization and Model analysis and other functions, model fine-tuning provides open source model library, model management, training task management and other functions, model deployment provides automatic deployment, operation monitoring, standard API and automatic packaging SDK, etc.
Model evaluation mainly serves the selection of open source large models, including testing with standard corpus or custom corpus, after conducting A/B tests on different open source large models or different versions of the same large model, and analyzing the relevant test results Perform analysis and visualization, and then select the large model to be fine-tuned based on model parameters, resource usage, etc.
(Example of model fine-tuning on Appen Intelligent LLM large-scale model development platform)
Model fine-tuning is to use high-quality labeled data and RLHF artificial feedback to enhance learning on the selected large model, and fine-tune it for different scenarios. The results of model fine-tuning will be returned to model evaluation, and the two will be linked to complete model iterations until the desired effect is achieved. Model deployment is to deploy the fine-tuned large model to the customer's computing resource environment, and can serve externally in the form of API or SDK.
The computing resource management of Appen Intelligent LLM large-scale model development platform is to manage tasks and schedule resources for customers’ computing resources, including CPU and GPU resources, as well as support and scheduling for upper-layer applications.
The Appen Smart LLM large model development platform can use the basic large model developed by Appen China itself, or the open source basic large model owned by the customer or a third party.
In terms of self-developed basic large models, the Appen China R&D team is mainly based on the work of the open source community, and is also evaluating other selection solutions horizontally. The feature of Appen China's self-developed large model is mainly fine-tuning on its own data set, including conversations on general topics and corpus with professional backgrounds. Appen itself provides more than 250 pre-labeled audio, image, text and video datasets. These high-quality labeled datasets are very precious for large-scale model pre-training. In addition, the Appen China R&D team is still paying attention to the progress of the academic and industrial circles, and constantly optimizes the self-developed large model in terms of model structure, optimization methods, and deployment efficiency.
In terms of cooperation with third-party large models, Appen Global has in-depth cooperation with NVIDIA, AWS, etc., especially with large models such as NVIDIA and enterprise-level AI development platforms. Combining data services, etc. with large-scale models, AI platforms and tools from major manufacturers, to provide industries and enterprises with end-to-end one-stop generative AI solutions. In addition, Appen also cooperates with enterprise-level large-scale model startups such as Cohere and Reka AI to provide highly secure customized proprietary models. In China, Appen China also cooperates with well-known basic large-scale models to understand the characteristics and applicable scenarios of these large-scale models, and provide customers with professional model selection solutions and consulting services.
Technology co-creation, growing together with AI leaders
As a provider with rich experience in delivering data, the biggest competitive advantage of Appen's intelligent LLM large model development platform lies in its rapid response to iterative models and data delivery from the perspective of zero-sample and semi-supervised learning.
As a long-term supervised learning data labeling service provider, Appen has a lot of practical experience in project delivery, and can continuously mine the value of data for LLM training and fine-tuning tasks as well as generative AI applications. Key advantages include:
First, grow together with the customer's algorithm application. When industry customers practice generative AI in the early stage, it is difficult to clarify the project requirements at the beginning, and they need to cooperate, explore and develop at the same time, and finally complete the application construction through continuous iteration.
Appen is good at managing and coordinating the delivery cycle. It can deliver data sampling, model optimization, application testing, etc. to customers in batches. Model optimization can be carried out alternately with data sampling; it can use small samples and incremental learning to drive models in Iterates quickly in the project, and the data collection standard is integrated into the application test faster; the labeling project can even be regarded as the "pre-quality inspection" before the customer's LLM performance test, which quite pre-empts the knowledge of the customer's industry or business scenario, and also It is pre-training pre-training.
Second, better grasp "Human in the loop". The development platform will analyze the various interactive behaviors of human beings in the process of collecting and labeling, and Appen has rich experience in this area, which can be converted into the "reward function" in the RLHF algorithm, and can mine more fine-grained labeling information Etc., providing more data nutrients for the preparation of large models, reflecting a deep understanding of the data mining dimension.
Third, long-term cooperation can bring scale effect in data collection and standardization. Appen's development platform has large-scale, safe, and high-quality data, as well as complete industry benchmark models for vision, text, and voice, and has been practiced in multiple standard acquisition projects. At the completion stage of each project, a model algorithm and high-quality data benchmark with good performance and complete mirroring of customer needs can be produced.
New Appen: A full-chain AI service provider
Since the LLM large model became popular around the world, there has been an argument that the large model will rule the AI world. However, after the "Hundred Models War" in the first half of 2023, everyone gradually realized that deep learning and LLM large models are equally important for AI applications. The so-called LLM large model refers to a general-purpose basic AI large model with model parameters of tens of billions or more than 100 billion, which has "intelligent emergence". However, due to the characteristics of large parameters and large computing resources, it is not suitable for enterprise and industry scenarios. Deep learning And machine learning has irreplaceable value in practical applications.
Dr. Tian Xiaopeng, Appen's Global Senior Vice President and General Manager of Greater China and North Asia, said that for the future AI market and smart economy, Appen's strategy is to use both deep learning and large models.
First of all, deep learning and machine learning are playing an effective role in the current digital transformation, especially for enterprise-level scenarios such as real-time computing and edge computing, and play an important role in the fields of smart cars, smart Internet of Things, and smart manufacturing. , Appen will still insist on AI data services for deep learning and machine learning, and at the same time form an end-to-end solution with the deep learning and machine learning platforms of top AI companies to meet the current needs of enterprises for AI engineering implementation.
At the end of June 2023, Appen's enterprise-level high-precision artificial intelligence-assisted data labeling platform - MatrixGo officially launched the SaaS version. Since its release, MatrixGo has experienced thousands of AI data labeling projects and has accumulated rich practical experience from various industries and various types of projects. The launch of the MatrixGo SaaS version allows enterprise customers to deploy MatrixGo more quickly, open and use it as soon as one day, and put it into production. At the same time, they can obtain professional training and customer service support. The SaaS version will continue to provide customers with the latest version of MatrixGo that is updated immediately. , allowing enterprise customers to use the latest and most advanced data services to create high-quality deep learning and machine learning applications.
In addition, Appen also uses LLM technology to improve data annotation tools and platforms, and continuously strengthens Appen's competitive advantage in deep learning and machine learning data services. The newly launched document intelligence product can automatically extract information from unstructured documents, such as extracting content from scanned documents or document photos, with an accuracy rate of 99%, which greatly expands the source of enterprise AI data. NLP automatic labeling uses small sample or zero sample learning and LLM model to automatically label data, thereby accelerating data supply. In 2022, Appen also invested in MindTech, the world's top visual AI synthetic data provider, which can provide a series of high-quality, multi-dimensional and multi-angle synthetic realistic pictures to deal with the problem of small samples or even zero samples.
Secondly, Appen will make strategic investment in the LLM large model and launch the LLM product line represented by the Appen intelligent LLM large model development platform. Appen's LLM product line includes four major parts: basic data, baseline model, evaluation and fine-tuning, and upper-level generative AI applications.
LLM basic data provides finished data sets, data crawling, data cleaning and open source data, etc., and provides high-quality data sets for LLM basic large model training and fine-tuning large models.
The baseline model provides self-developed models and third-party open source or commercial models, and supports customer-owned models. Appen's self-developed models can be customized according to usage scenarios, and the privatization volume of models can be limited according to operating resource requirements, supporting privatized deployment , cloud platform API calls, etc., third-party cooperation models include Reka, Cohere and other excellent commercial and open source models at home and abroad.
Evaluation and fine-tuning include LLM large-scale model training services such as expert corpus, RLHF, A/B testing, and model evaluation. Appen has millions of crowdsourced data collectors and labelers around the world, supporting 235+ languages and dialects, as well as expert crowdsourced resources for financial, retail, industrial and medical industries. In the past, these resources served the data annotation of deep learning and machine learning; in the future, facing the training needs of LLM large models, these resources can also provide prompt words-output corpus, professional field corpus, and artificial embedding into LLM large models The artificial feedback enhancement link of the training implements the RLHF algorithm and improves the professional field capabilities of the model.
Model evaluation includes methods such as A/B testing, model evaluation, red-blue confrontation, and benchmark testing. Appen's LLM experts and crowdsourcing resources work together to evaluate the output results of different large models and different versions of the same large model. The output is evaluated to avoid risks such as discrimination and pornography, to evaluate the ability of the model in multiple rounds of dialogue confrontation, and to benchmark large models using industry standard corpora.
Third, in the longer term, Appen will combine deep learning and machine learning with LLM large models to develop end-to-end generative AI applications for enterprise customers, providing full-chain consulting from data to models to application development and application development services, and then become a core AI supplier.
Compared with other participants in the LLM and generative AI track, Appen has a very solid data "chassis" and a full chain of data tool chains, platforms and human resources, and data capabilities are the king of LLM and generative AI . In addition, Appen has a 27-year history of cooperation with global AI companies and AI ecosystems. It has also participated in a large number of enterprise and industry AI implementation projects, and has rich experience in implementing enterprise-level projects. These have laid a solid foundation for Appen's self-subversion in the era of LLM and generative AI.
Looking forward to the future : LLM large models and generative AI are the "singularity" of global intelligence evolution, and a new attitude of Appen is emerging from LLM large models and generative AI. From a top AI data service provider, to quickly entering the industry LLM large-scale model and generative AI track, and then to generative AI applications and full-chain AI consulting development, Appen is based on the accumulation of the past 27 years, in the global intelligent evolution " "Singularity" moment, seize the opportunity, change quickly, and work with the global AI ecosystem to open up the great future of generative AI! (Text/Ning Chuan)