AIGC is quietly becoming a necessity in all walks of life.
@商科星球original
Author丨Yuan Jing
Edit丨Big Rabbit
The domestic large-scale melee has lasted for more than half a year, and traditional Internet giants and artificial intelligence companies have entered the scene one after another. At the critical moment of the "Hundred Models War", the industry is quietly splitting. Some more forward-looking enterprises began to look into the distance, preparing for an "asymmetric" business competition.
At the closed 2023 World Artificial Intelligence Conference, Shanghai Tianrang Intelligent Technology released a "three-piece set" of "Tianrang Xiaobai" large-scale model products, including a large-scale general language model, an application development platform and a semantic search engine. Among them, the parameter scale of Tianjie Xiaobai language large model reaches 186 billion.
Since its establishment in 2018, T&N has been focusing on the research of general artificial intelligence. At present, the company mainly serves digital scenarios such as enterprise services, digital finance, biotechnology, and urban operations. During an in-depth exchange with Digital Planet (ID: digital-planet), Xue Ronggui, founder and CEO of Skyland, believes that AIGC technology has become a new generation of "water, electricity and coal", and the emergence of a large number of artificial intelligence companies will reshape the future business structure.
01
The Potential of Generative Content Technologies
In the process leading to the AGI era, large language models are playing an extremely important role (although AIGC technology includes but not limited to large language models, the latter is very important). Logically, the large language model is the premise to leverage the interaction between human and machine, and has been fully applied in text generation, machine translation and dialogue systems.
It is conceivable that after the image generation, speech generation and 3D engine are continuously improved, the application fields and capabilities of AIGC will continue to expand and improve.
Among the many large-scale model products, the advantage of Tiandi Xiaobai lies in multiple rounds of dialogue and logical reasoning.
In the large-scale model industry, multiple rounds of dialogue help to understand the intention of the current dialogue and the logical relationship behind the intention, so as to generate more accurate answers. This can be seen as the key to more intelligent interactions and services. Technically, Tiandixiaobai can establish long-term memory and context perception of language, so as to better respond to human's natural language communication and expression needs.
In Xue Guirong's understanding, the logical reasoning of large models is changing the rules of the industry. "In the past, we have also tried small models, but in the context of prompt words, the results obtained by a model with a limited range of knowledge are not rich." In addition, the support of small models for multiple rounds of dialogue is also relatively limited.
In addition, Tiandixiaobai also combines the thinking chain technology. Technically, it can decompose a complex task into multiple subtasks and process these subtasks in parallel at different levels, thereby improving the efficiency and accuracy of the entire task. .
Specifically, when a large-scale deep learning model needs to complete a complex task, it is usually designed to consist of multiple sub-modules, each of which is responsible for processing a specific sub-task. These subtasks can be image classification, object detection, speech recognition, etc.
During inference, these subtasks are handled independently, and each submodule is analyzed using different data and parameters. When all tasks are processed, it can combine these sub-modules to form a complete task solution.
Technicians inside Tianrang believe that this chain of thought technology can help dismantle complex tasks and break them down into manageable sub-tasks. This makes the training process of the whole task more efficient, and knowledge and experience can be shared among multiple subtasks, thus improving the performance and accuracy of the whole model.
"In short, the comprehension, reasoning ability and analysis ability of large models are very different." He concluded. And this is also the reason why T&N continues to promote the large-scale model business.
02
Application of AIGC in Intelligent Transportation
As we all know, the emergence of AlphaGo is an important milestone in the field of reinforcement learning, and its success shows the great potential of reinforcement learning in solving complex problems. It gradually improves its level by playing against itself continuously, so as to make optimal decisions in complex Go games. This not only caused a sensation in the surrounding field, but also further opened up new possibilities for artificial intelligence to empower all walks of life.
Inspired by this, Xue Guirong and the T&R team ignited their confidence in AI. They developed their own AI Go TRgo, using only 1% of the computing resources of AlphaGo, and successfully defeated the world Go champion Park Tinghuan, becoming one of the first domestic teams to explore reinforcement learning. .
Not only has it made some gains in the field of Go, but in 2020, Tianrang will make important progress in the transportation industry. TRL launched TRTraffic, a city-level traffic congestion management system, and helped Nanchang become the first city in the country with "unlimited traffic".
However, on the way of exploring general artificial intelligence and training large models, Tianrang is also facing the problem of lack of data. Just like today's Chinese version of ChatGPT, high-quality Chinese corpus data is very scarce. In this regard, T&N seeks innovations, using robots to fight against generated data, iteratively optimizing models for survival of the fittest, and using reinforcement learning for data enhancement.
These methods have greatly improved the performance of the model, and also proved the superior performance of reinforcement learning in dealing with different challenges.
Now, with the further breakthrough of AI technology, he is more optimistic about the progress of AI in perception, cognition and decision-making. "We can use artificial intelligence to solve applications in complex scenarios." Xue Guirong said to Shuke Planet (ID: digital-planet).
An internal technical staff of Skyland said: "Artificial intelligence can help cities develop into unlimited cities, and deep learning and reinforcement learning technologies can be used in it."
03
Application of large models in various industries
In addition to making breakthroughs in large language models, Tiankong has also applied leading AIGC technology to the field of biopharmaceuticals, bringing great potential and opportunities to this field.
In principle, AIGC can generate completely new proteins, and even the generated proteins can be completely different from natural proteins in nature.
As a result, the new technology greatly shortens the time and cost of protein research, and opens up a new paradigm of computational biology research. At present, the applied technology can avoid the tediousness and uncertainty in the traditional protein design method, and can also greatly improve the quality and stability of the protein. Xue Guirong said to Digital Planet (ID: digital-planet): "The biotechnology industry will further explode, and AIGC will become the technological foundation for the next round of explosion."
In 2021, the company released the protein structure prediction model TRFold2. The prediction accuracy based on the CASP14 test set is comparable to that of AlphaFold2. In 2022, the company released the protein design model TRDesign to realize on-demand protein design. Specific projects include:
Published the protein complex structure prediction model TRComplex;
Released the orphan protein structure prediction model TRFold-Single;
Build xCREATOR, the first protein design workbench in China.
Rather than saying that "discriminative AI" makes judgments based on data, generative AI expands the boundaries of AI's thinking when making decisions because of its contextual thinking ability.
Macroscopically, the computing power, algorithms, middleware technology and data required by AIGC will stimulate a new round of investment boom; in the industry, all walks of life combined with AIGC will accelerate the popularization of intelligence; microscopically, embracing intelligence will speed up Faster companies can gain new market competitiveness.
04
The value behind the open platform
Although a large model is good, not any company can do it. Among them, one of the main reasons is that it is extremely challenging to build a growth flywheel that integrates users-data-capabilities:
First, diverse user needs and feedback can help large models to be continuously optimized and improved. The iteration and improvement of the large model requires the needs and feedback of different users as a guide. Users can use the large model to discover its existing problems and deficiencies, and provide feedback and suggestions to the model developer. These feedbacks and suggestions can help model developers better understand user needs, and improve and optimize the model in a targeted manner, thereby improving the performance and accuracy of the model;
Second, data is the basis for large model iterations and the emergence of intelligence. Large models require a large amount of data for training and optimization. Only with enough data can the model be continuously iterated and improved. At the same time, data can also help the model to better learn and understand natural language and knowledge, thereby improving the intelligence level and performance of the model;
Based on the above two points, it can be seen that the improvement of large model capabilities is a two-wheel effect composed of real user feedback and high-quality data iteration. The two are indispensable, and jointly promote the emergence and development of large-scale model intelligence.
Xue Guirong believes that the iteration of ChatGPT3.5-4.0 is a model of the above logic. In order to catch up with the foreign advanced level faster, T&R has formulated an open platform strategy to meet the challenges.
On July 7th, Tianrang released the open platform of "Tianrang Xiaobai". The platform has a built-in self-developed large language model, and integrates a complete set of tools and resources into the platform. It is understood that the platform aims to help developers easily create AI applications, explore large models more flexibly, and create AI innovative products with better experience and beyond expectations.
In view of the key bottlenecks in the implementation of large models at present: the "illusion" of large models, that is, serious nonsense, the security of private data, the lack of enterprise-level Chinese large models in the market, and the closedness of large models themselves, T&N launched " Three-piece set": that is, a large language model, a semantic search engine and a development platform. These three constitute a full-stack support system for developing smart applications, providing developers with a one-stop solution.
The first is the large-scale general language model, which is trained through deep learning with 186 billion parameters. It has ChatGPT-like capabilities and can perform core functions such as multilingual dialogue interaction, knowledge question and answer, and logical reasoning. It understands complex contextual information and responds precisely based on previous conversations.
Secondly, it is a semantic search engine, which further strengthens the ability to generate large models. Before and after text is generated, the engine retrieves relevant factual information and validates input and output to ensure that the content is correct, reasonable, complete, and consistent, and that large models do not “hallucinate” or provide incorrect or meaningless answers. In the actual test, the recall rate of Tiandixiaobai's semantic search model has a very good performance. Among them, in the comparison test with the recall rate of OpenAI's semantic search engines TOP1 and TOP3 (respectively counting the probability that the recalled TOP1 and TOP3 results contain the correct answer), Tiandi Xiaobai's results were 77% and 86%, both exceeding OpenAI 73% and 85% performance.
Finally, there is the application development platform, which adopts a visual interface, and what you see is what you get. After the user enters the preset prompt word (prompt) in the left column of the platform, the application effect can be tested in the right column immediately to ensure that the application effect meets the requirements. expected. The entire creation process takes no more than ten minutes.
At present, the open platform has greatly lowered the threshold for users to use, even if there is no programming foundation, application configuration can be realized through the visual operation interface. According to the internal staff of the open platform, developers can choose multiple model services with different parameters and precision, configure prompt word parameters and interaction types, and use the document set function for semantic retrieval according to different scenarios and needs, so as to help users create more than expected AI application, so that the application effect reaches the best state.
Users can also choose to call other large models or enterprise private models according to specific needs.
Ending: In the in-depth dialogue, Digital Planet (ID: digital-planet) heard a number of entrepreneurs admit that OpenAI has a certain degree of first-mover advantage.
But in China, with data precipitation and reinforcement learning technology innovation, the iterative effect of large models is showing.
In the words of Xue Guirong: "Only by mastering more high-quality data can we get closer to human intelligence." Now, the breakthrough point of Tianlang is to continue to increase investment in large models to achieve full coverage of scale, depth, and breadth.
As a technology-based company, Tianlang has initially demonstrated its potential to challenge AlphaFold, and has made cutting-edge breakthroughs in AI Go, intelligent transportation, biotechnology and other fields.
Time is on the side of Chinese companies, perhaps, in the future we can witness the emergence of more outstanding companies. Now, we are stepping towards the age of intelligence.