Smart stuff
Author |Cheng Qian
Edit |Desert Shadow
The fire of large models has burned to personal terminals, especially personal computers related to our life and work.
Entering 2024, the focus of competition in the large model industry has extended from volume parameters and volume performance to commercial implementation, and various AI applications and tools have burst out, allowing users to have a stronger perception of large models.
On the one hand, some enterprises continue to run fast on the road of improving the capabilities of general large models, and on the other hand, some enterprises have begun to explore the possibility of deploying large models on terminal devices. Mobile phone players from all walks of life have deployed large-scale model technology, and the intelligent assistant experience has been subverted; The self-developed large model of the rim is put on the car in batches, redefining the intelligent cockpit experience; There are also manufacturers of smart home, robots, chips and other tracks are also accelerating their entry into the ......
As a super productivity tool in various industries such as medical care, education, and manufacturing, PC is likely to set off a revolution in productivity tools after plugging in the wings of large models.
Just yesterday, Lenovo Group, a leader in the PC industry, released its financial results for the third quarter of fiscal year 2023 24. According to the financial report, Lenovo Group's revenue in the third fiscal quarter was 113.6 billion yuan, a year-on-year increase of 3%, and the net profit was nearly 2.6 billion yuan, and its profitability increased quarter-on-quarter for two consecutive quarters. Then the company released a hard-core science popularization ** at the same time, revealing the originality of Lenovo AI PC.
*Through cool techniques, the AI PC was explained interlockingly, and the AI OS ecology and AI chip TAC developed by Lenovo were realistically presented with three-dimensional animation, stereoscopic storyboard, plane storyboard, etc., and cutting-edge technologies such as large model compression, knowledge distillation, and vectorization were clearly popularized.
In 2024, the wave of productivity tool revolution is coming, ushering in the era of AI for You, and Lenovo Group is exploring a new possibility for the AI PC industry.
From the initial appearance of chatbots, to now the products and tools built on the capabilities of large models have gradually transformed into more practical productivity tools.
While individual users integrate these products into their daily life and work, their needs are also more in-depth and subdivided, and they need personal models that are more in line with their own usage habits.
Many of these demand-side trends are pushing back the transformation of the large model industry. Large models are no longer limited to ultra-large-scale parametric theory, and compression technology is used to reduce the size and computational complexity of the model while maintaining stable performance, making it easier to deploy.
When the large model can be deployed to the device, a question arises about how to make the capabilities of the large model more suitable for the user's individual needs.
PC devices stand out from the crowd and have advantages in terms of interaction mode, terminal computing power, and local storage capabilities, making this terminal device an important way for AI technology to benefit everyone.
First of all, the large model drives the transformation of human-computer interaction, which can achieve high-quality recognition and understanding of voice, text, image, and **, so as to make the way of human-computer interaction more natural, intelligent and efficient. At the same time, PCs can connect to more external terminals, which is the core of a variety of interaction methods, and the experience will be richer in the future.
Secondly, PC is more suitable for the presentation of the corresponding creative ability of large models, from the original large language model to the multimodal large model, the content generated by it has been extended from simple text to **, audio and even 3D animation.
Finally, there is computing power and data. PC is the strongest computing power product in personal terminal devices, and it is also more suitable for users to store personal data and files, so it can support large-scale model computing and protect user data security.
In summary, it can be seen that the characteristics of PC and large models have natural adaptability.
The advent of AI PCs is further amplifying the productivity tool attributes of PCs.
First of all, we must answer the most essential question: what is an AI PC? Yang Yuanqing, chairman and CEO of Lenovo Group, gave five characteristics of AI PC, compared with traditional PC, AI PC has stronger computing power, larger storage space, large model with reduced volume but no reduction in performance, complete AI application ecology, security and privacy protection.
The first three are related to the basic hardware and software infrastructure, that is, the ability to run compressed and performance-optimized personal large models on PCs, heterogeneous computing including CPUs, GPUs, and NPUs, and the local knowledge base that contains users' personal data.
The latter two test the ecological construction capabilities of PC manufacturers, so that AI PCs can have richer forms of natural language interaction, not only text, voice, and gesture interaction. The last is to make AI PCs more reliable and make users' data safer.
Based on this, in the future, PCs can not only carry the ability of general large models, but also independently run personal large models exclusive to users, becoming a more practical tool. Behind this, there are many challenges.
The AI PC should not only compress the ultra-large-scale parameter model, but also ensure the availability of the PC and maintain stable performance.
On top of the compressed large model, the AI PC needs to build its own personal large model for users, maintain continuous learning and training of user habits, data and other content, and have powerful computing power and storage space to support the training and storage of these data.
More importantly, personal data privacy and security issues are closely connected with users' daily life and work, so the requirements for user data security protection will also be improved by leaps and bounds.
And now, with these technical difficulties overcome, AI PCs will revolutionize productivity tools.
How much black technology is hidden in the AI PC displayed by Lenovo Group, which is only less than 2cm thick?
*Using 3D animation, the technical points behind the ingenuity of the AI PC are restored with modeling effects, and the AI PC is disassembled in three dimensions, we can see that the foundation of the AI PC is powerful heterogeneous computing capabilities, and the high-performance CPU, GPU, and NPU form the preliminary hardware architecture of the large model.
On top of the hardware architecture, how to achieve a large model that can be controlled by a PC? In fact, it is difficult to install a model with a large parameter scale into a PC, and when quantifying it with a set of numbers, we found that the large model GPT-3 previously released by OpenAI5 The parameter size is 175 billion, and it is expected to occupy more than 330GB of memory.
The answer to this problem is knowledge distillation, in which the model technology of knowledge distillation is visualized in the form of topological diagram through the technique of plane storyboarding.
It can be seen that knowledge distillation is to use the supervised information of the large model to train a lightweight small model, so that the model can obtain better performance and accuracy with a small parameter scale. Among them, the mature model is the "teacher", and the local model deployed in the PC is the "student".
In this way, the parameter size of the large model can be compressed from 175 billion to 7 billion, and the personal large model only needs to occupy 4 to 5 GB of space in the hard disk.
Now that the big model in the PC is ready, the next step is how to make it more user-friendly and tailor the most powerful productivity tools for everyone.
*The use of stereoscopic storyboards to demonstrate how to build a local knowledge base through vectorization technology, when the user authorizes the insertion, personal documents, ppt, pdf, schedules and other information related to personal habits will be stored in the local knowledge base.
This knowledge is vectorized into a form that can be computed by AI and then stored. When a user puts forward a requirement, the personal model will also vectorize the corresponding requirement, and query and match the most similar fragment in the local knowledge base. After the personal model extracts the relevant text blocks, it will output the answer according to the user's prompt words, combined with the generation and comprehension ability of the large model.
In addition to the deployment of intelligent computing power and the local knowledge base of personal large models, the AI OS independently developed by Lenovo Group is the ecological entrance of the entire computer, which can call local documents and regulate hundreds of hidden systems and hardware settings to complete more personalized needs of users.
Lenovo Group has broken the pattern that AI knowledge in the field is only dominated by models, allowing AI to return from "general" to "individual" in the PC field, and truly realize "thousands of people, thousands of faces".
When a user interacts with an AI OS, for example, "What should I do if I feel that the computer is not smooth enough to play the game?" AI OS will identify the keywords involved in this requirement, and then understand the relevant settings that the user wants to adapt to the game scene, and pop up the setting option of "Turn on Beast Mode".
With the increasing number of interactions triggered by users on PCs, AI OS has a deeper understanding of individuals, and can solve many pain points in users' daily life and work, such as generating meeting minutes and retrieving past PPTs, so that they can serve individuals and bring customized and efficient experiences to users.
It is worth noting that every response and learning of a large model will occupy a large amount of system power, which requires AI PC to intelligently schedule and allocate system power to make the user's experience better.
For these problems, Lenovo Group's self-developed AI chip TAC (Trusted AI Controller) is the key, which shows the strength of TAC to assist in scheduling computing power allocation and release high performance through gorgeous and smooth three-dimensional animation.
TAC connects with all the hardware on the PC and transmits the state of that hardware to the AI OS. As shown in **, these states will be understood by AI OS, and they will be identified as specific use scenarios such as meetings, games, and learning.
Based on different scenarios and task identification, TAC will accurately divert the energy consumption pressure to ensure the reasonable operation of computing chips such as CPU, GPU, and NPU to ensure the sustainable and efficient operation of AI PCs.
This chip will not only support the maximization of efficiency and power consumption, but also become a powerful guard for the protection of users' personal data.
TAC will encrypt the user's personal files, usage habits and other information through the key, and seal the key in the ROM of the chip for independent power supply, effectively preventing data leakage and malicious attacks.
Finally, Lenovo AI PC ingeniously combines software, hardware and ecology to provide a landing solution that makes it realize it on a personal terminal for the first time, and layers of black technology make us see that the personal AI twin (Personal AI Twin) is about to become a reality.
AI PCs are booming and about to land. AI will no longer be a topic of conversation for users, but a tool that can be used.
As mentioned in Lenovo Group's popular science**, after PC stepped onto the stage of history 39 years ago, it has carried almost every technological change in the digital world, and today, 39 years later, AI PC may also be about to become one of the best solutions for human interaction with the digital world.
This time the AI PC is both a breakthrough and a return. Facing a moment of major changes in productivity tools, Lenovo Group, as a leader in the PC industry, has brought the disruptive power of artificial intelligence to every user's terminal, and let everyone realize AI inclusion with AI PC.