Perhaps we have misunderstood AIPC.
AIPC, which stands for AI and PC, appears to be a crude combination of AI and PC. The most common interpretation is that it is a product of large models being integrated into computers, requiring a reduction in model size while protecting privacy.
It was proposed by leading players like Lenovo in 2023, and corresponding new products are set to be mass-produced and showcased at CES 2024. This speed reflects the terrifyingly high efficiency of today’s PC supply chain, while also revealing the anxiety throughout the entire industry chain—declining PC shipments, the consensus that larger models are more capable, the interdependence between models and the cloud in this wave of AI, and the public’s eager expectation for new AI hardware, all deepen this impression.
However, all of this is actually based on a seemingly established “consensus” today: we have entered an era dominated and determined by large models, where models will dictate the future of all software and interactions. Models are consuming the world; any artificial intelligence that does not start from models is merely a reactive response.
But is this consensus really the direction of the future?
Some frontline projects that are striving to unleash model capabilities are presenting a different picture. For example, early attempts at agents like AutoAgent initially focused on developing all functionalities around models, but more and more issues arose. It later began to incorporate tool-using capabilities, which truly started to reflect the characteristics of an intelligent agent. For instance, hardware companies like Apple are shifting their research focus to hardware and computer architecture, integrating models into existing features such as hardware, storage, and operating systems. Moreover, there is a renewed interest in how RNNs can help with model complexity, and the RAG design approach, which utilizes “outdated” search technologies, is gaining popularity… Essentially, they all leverage “previous generation” technologies, and the “consensus” around models is beginning to loosen.
These all point to another route: models are not everything; they are an important part of the composition of intelligence, but they require other technologies rather than replacing all technologies. Models do not equal intelligence; agents are not a branch application of models, but rather a new intelligent system that encompasses models.
This represents two different judgments about the future of AI. Currently, no products have reached either of these ideal states, but how we think about this issue today will still determine the path a company and its products take towards future transformation.
I believe that understanding AIPC should start from the second route, as it has enormous potential. At CES, I had the opportunity to discuss this thought with senior executives from Lenovo. This company, which ranks first in PC shipments, is also trying to define what AIPC will be in the next era.
“We are not chasing AI in the short term; this has long been our long-term strategy,” said Yang Yuanqing, Chairman and CEO of Lenovo Group, during CES 2024. As the earliest proposer of the AIPC concept, Lenovo unveiled more than ten prototype AIPC products at the conference, including the Yoga Pro 9i, which provides AI creative tools in the Yoga Creator Zone, the ThinkBook Plus Gen 5 Hybrid, which seamlessly switches between laptop and tablet modes, the business AI PC ThinkPad X1 Carbon AI, and the new generation ultra-compact ThinkCentre neo Ultra. Additionally, Lenovo plans to introduce AI now, a large model assistant software, in some of its products in China later on.
Lenovo claims these products are currently “AI ready.” From these releases, it is evident that Lenovo has actually done several things at this stage: systematically optimizing models based on understanding of hardware, designing prototypes for AI software and user interaction, and providing interfaces with more possibilities.
“Currently, these products can only be called AI-ready products, primarily utilizing graphics cards, whether integrated or discrete, to enhance their computing power. The PCs we see now can achieve about 10 Tops of computing power, while we believe the ideal situation is to reach 40 Tops before we can officially launch the first generation of AIPC later this year,” Yang Yuanqing said.
Yang Yuanqing summarized the ideal characteristics of AIPC in five points: firstly, it must have a large model, being driven by a large model AI; secondly, it requires a powerful and efficient heterogeneous computing hardware and software platform; thirdly, it should have self-updating and progressive capabilities, possibly achieved through robust storage; fourthly, it should possess a natural interaction capability that resembles human-to-human interaction; and fifthly, it must have strong privacy protection capabilities.
“To move towards intelligent agents, three capabilities need to improve: first, the ability to interact with and perceive the external world; second, gradually developing the ability to understand and utilize tools; and third, having long-term and short-term memory for planning,” said Dr. Rui Yong, Senior Vice President and Chief Technology Officer of Lenovo Group.
This is an interesting judgment. When you consider these three factors in relation to computer architecture, you will discover something intriguing: PCs possess all the “basics”—as a complete hardware system, their camera sensors and expandable plugins can better perceive the physical world; compared to other intelligent agents, their environment naturally contains various tools—those rich applications, especially productivity tools, serve as a ready toolbox for an agent; and the storage architecture in computer systems is frequently referenced in designs for long and short-term memory for large models.
So, how should we now understand what AIPC is?
AIPC is a more powerful AI agent, and this agent is not what many people have previously understood as the “application form of models” under the guidance of OpenAI, but rather a new form of AI that transcends models, a form that is closer to AGI.
“I understand the five layers of the world model; the fourth layer is a single agent, and the fifth layer is possibly multiple agents. Humans are multiple agents, and achieving that may lead to AGI,” he said when I shared this viewpoint with Rui Yong.
In the process of realizing this super agent, all participants are no longer merely cutting corners but are leveraging their expertise in system and hardware optimization, utilizing all existing technological solutions and architectures, as well as future emerging technologies, to fill in the remaining gaps for the development of this agent.
According to Rui Yong, Lenovo has made many optimizations on KV Cache to enhance the contextual memory capability of edge models through hardware and system-level optimizations, thereby avoiding excessive spatial occupation when increasing context length algorithmically. Additionally, as a model that needs to operate offline, it must also be a multimodal model.
“To achieve this, one must have a deep understanding of computer architecture, as sometimes the bottleneck is not in computation but in memory bandwidth, and secondly, they must have a solid understanding of the algorithms behind models like GPT. For instance, it outputs one token at a time, adding it to the input, thus increasing the input length. However, during computation, most operations are multiplication and addition, so a sufficiently smart person can realize that many of these have already been computed and find ways to determine which previously computed values are likely to be reused and which are not,” he explained.
What PC manufacturers have been doing is optimizing algorithms and forming product solutions through hardware and software integration, and Lenovo’s own LA series AI chips are also applied to optimize CPU resources, storage resources, and performance in model training and inference behind these released products.
“All these aspects require significant effort; missing any one of them is not acceptable,” Yang Yuanqing said.
A well-rounded AIPC is closer to the form of an intelligent agent than a standalone large model, and PC manufacturers are closer to AIPC, bringing great opportunities for them.
Welcome to join this planet and witness the development of the silicon-based era ↓
Give a click on “View” and then leave👀
