In the ever-evolving field of artificial intelligence, a thought-provoking equation has emerged: 80% data plus 20% models equals better AIThis assertion is brought to us by Andrew Ng, a renowned computer science professor at Stanford University, who famously articulated this principle on his 45th birthdayHe emphasized the crucial role data quality plays in AI development, stating that if 80% of the work focuses on data preparation, then ensuring that data is of the highest quality becomes a vital concern for AI teams.

Transitioning this principle to the realm of autonomous driving, we see a similar need for a consistent and reliable exchange of high-quality dataThe pursuit of creating an effective intelligent driving system appears enticing, but navigating the complexities of data infrastructure, computational horsepower, and algorithmic execution presents significant challenges for automotive manufacturers

High-quality data and ample computational resources are prerequisites for every technical upgrade; without these foundational elements, even the most sophisticated algorithms cannot reach their full potential.

As the industry grapples with a pivotal phase in the competition for intelligent driving, the focus has shifted from merely recruiting top engineering talent to refining model design, robust toolchain development, and rigorous testing protocolsA conspicuous observation is that, whereas previously, team size was a key metric of advancement, by the latter half of 2024, the primary competitive measures will revolve around cloud computing power and the quality of data.

From the initiation of large-scale deployments in early 2023 to the more nuanced battles over "space allocation to vehicle" projections in 2024, several pressing questions have surfaced: How can data loops be effectively established? What steps can companies take to create efficient reservoirs of computational power? How can companies maximize data and computational resources while achieving optimal algorithms? These queries are fundamental to automotive manufacturers striving to secure a competitive edge in the sphere of intelligent driving.

Beneath the surface of this ongoing technological development lies the concept of data closure or feedback loops

• The Evolution and Constancy of PCs
• Commodity Prices Fall as the U.S. Enters Recession
• Time to Dive into AI Glasses Before It's Too Late
• Bank Collapse and Plummeting Stock Prices
• Impact of Significant US Dollar Appreciation

The intelligent driving ecosystem operates under an end-to-end methodology where data collection, storage, analysis, and annotation are integral stages that feed into model training, simulation validation, and deploymentEssentially, creating effective data loops pertains to harnessing the value within substantial amounts of driving data, gradually transferring human driving knowledge into the parameters of autonomous modelsThis intricate process substantially enhances the human-like performance of intelligent driving systems, ultimately leading to safer and more natural driving experiences.

Through a developmental lens, intelligent driving technology initially focused on hardware-driven approaches and a nascent understanding of the data loop conceptAs the technology matured, the significant roles of algorithms and software began to shine, leading to the adoption of smaller models grounded in rule-based systems to tackle complex driving challenges

Today, in light of the increasing sophistication of intelligent driving capabilities, the reliance on data-driven methodologies has prompted the industry to move towards the third stage, where extensive data loops are paramount.

Companies like Tesla have set specific thresholds, stating that an effective end-to-end autonomous driving model necessitates at least one million high-quality video segments displaying diverse driving environmentsWith ten million segments, the system's capabilities could become almost extraordinarySimilarly, automotive manufacturer Ideal announced it would deploy end-to-end models trained on a staggering ten million segments by early this year, while rival firm XPeng claims to have amassed over twenty million training video clips.

The data acquisition landscape currently offers two primary avenues for intelligent driving firms: First, mining data from mass-produced vehicles

For example, for each model sold, engineers can implement specific rules; if a user’s driving behavior meets designated criteria, specific data (after anonymization) will be uploadedUsers can also voluntarily contribute atypical driving casesSecond, companies often comb through legacy dataEarly in the development of intelligent driving, a vast accumulation of data—much of it deemed ineffective—led engineers to rely on algorithmic rule sets to distill valuable information.

The importance of high-quality data akin to nutrients represents a defining line in the evolution of autonomous driving systems, continuously testing automotive brands' capacity for closed-loop automation in intelligent drivingRecently, with the scaling of mass-produced vehicles, many manufacturers have transitioned to collecting data through a shadow mode that emulates real-world conditionsThis emerging approach, however, remains fraught with challenges.

Foremost among these is the question of data collection strategy

How can firms balance the long-tail problem of data—assuring effectiveness—against the cost implications of aggregating vast datasets? Overly lenient data collection efforts often lead to a swamp of irrelevant data, whereas stringent criteria may result in the loss of valuable insightsSecondly, there exists the matter of data qualityDefining what constitutes ‘good data’ can be a complex challenge, as "bad data" (characterized by poor driving practices or rule violations) can significantly undermine training efficacyThirdly, managing data distributions—effectively evaluating features, understanding statistical distributions, and considering various dimensions—requires substantial dedication and intentional effort.

Automotive manufacturers face yet another challenge: the risk of insufficient data generalizationVaried configurations among different vehicle models often render datasets incompatible or ineffective across platforms

With numerous domestic car models in service, there is a real possibility that vast caches of data can ultimately become liabilities instead of assets—merely contributing to storage costsIt is not an exaggeration to state that data can account for over 80% of the overall research and development expenses in autonomous driving system development.

Therefore, the sooner manufacturers can build effective data loops, the better positioned they will be to create comprehensive and sustainable technological barriersThis establishes a framework for maintaining a competitive edge while keeping contenders at bay.

The escalating “arms race” for computational power catalyzed by AI large model paradigms has now prominently entered the automotive sphereNew age companies such as Li Auto, Huawei, and XPeng are pushing the boundaries aggressively.

Following trends seen with AI large models, the end-to-end autonomous driving technology—boasting billions of parameters—is also gaining ground towards hundreds of billions

The fierce competition for computational resources has emerged as a new cornerstone in the intelligent driving domain, now rivalling the significance of data acquisition.

This competitive arena has manifested primarily due to the escalating demands on data collection and processing capabilitiesAdvanced intelligent driving systems have proliferated, leading to an increase in the diversity and quantity of onboard sensorsIntelligent driving systems must continuously collect and consolidate data from these sensors to make informed decisions in real-time regarding navigation and control functionsThe rapidly growing demands for vast data processing and ultra-low latency create a burgeoning demand for computational power, which has risen exponentially.

It is generally understood that with each elevated level of automation in vehicle capabilities, the required chip processing performance increases tenfold

Intel estimates that Level 5 fully autonomous vehicles will require around 4,000 gigabytes of data to be processed every secondAdditionally, the advancement of smart cabin technology and vehicle connectivity has propelled new demands for computationUpgraded interaction models within cabins rely increasingly on screens, and the proliferation of in-car entertainment services enhances operational complexity—further necessitating higher computational capabilities.

Last year, Li Auto reported its cloud computing power at 2.4 exaflops, with a subsequent increase to 6.83 exaflops by NovemberXPeng plans to escalate its cloud power from the current 2.51 exaflops to a target of 10 exaflops by 2025. Remarkably, Huawei's intelligent driving division expanded its cloud power from 5 exaflops to 7.5 exaflops within just two months.

Currently, the bulk of automotive computing power stems from onboard platforms, which ultimately set the ceiling for how much software can be integrated into vehicles, thereby dictating the total lifecycle value of each vehicle

As software technologies are poised for continued improvement, manufacturers are proactively embedding significant computational resources into hardware to maximize future software service revenue collection.

However, according to Moore's Law, the potential of onboard computing platforms will inevitably plateauThe inherent characteristics of automobiles limit the extent to which manufacturers can indefinitely increase hardwareThis insatiable thirst for instantaneous data processing consumes ever-increasing computational demand and cultivates an environment wherein the race for computation becomes an unyielding and nonsensical competition, thereby inciting manufacturers' anxieties over computational sufficiency.

To alleviate the tight supply of computational power, cloud-based large models present a viable solutionWithin current market trends for developing end-to-end technologies, a trio of methods have emerged: one involves aggregating many rules and small models into a larger model, which demands numerous skilled rule engineers; two focus on deploying large models directly within vehicles for immediacy, albeit constrained by onsite computational capabilities; and three revolve around cloud-based large models (Foundation Models) that boast vastly higher parameter volumes compared to their onboard counterparts

However, training such cloud-based models necessitates incredibly robust computational and data processing capabilities.

The myriad challenges include enhancing concurrent training efficiency, ensuring reliable training durability, managing vast multi-modal data processing capabilities, and minimizing storage costs while maintaining data performance standards.

From a holistic vehicle perspective, as the edge and cloud computing frameworks align, the next competitive frontier will center on data mining capabilities, efficiently harnessing data, the entire technical stack's comprehension of available data, and striking a balance between computational efficiency amidst expansive infrastructure.

Essentially, the objective remains to capture high-quality, vast datasets and marry them with superior computational resourcesThis duo will empower the iterative enhancement of algorithmic capabilities, fostering a robust data loop across vehicle, cloud, and road systems, yielding efficient, cost-effective operation.

The foundation of intelligent and connected vehicles lies in their AI backbone

Whether exploring intelligent driving, smart cockpits, or integrated road-cloud systems, these innovations continue to evolve in alignment with burgeoning AI capabilitiesThe integration of deep learning and large models is progressively embedded within these technological frameworks.

For intelligent vehicles, the progression toward higher levels of automation is intricately linked to real-time environmental perception and analytical capabilitiesAmple, quality datasets and robust computational resources underpin these needsWithin the confines of resource usability, a vehicle-road-cloud network integrating communication, perception, and computation holds promise as a means for increasing the efficiency and cost-effectiveness of intelligent driving solutions.

This network fosters established feedback loops extending across vehicles, roadside data collection points, and cloud processing capabilities

Roadside data can supplement the inherent limitations faced by vehicles—such as blind spots or reduced detection capabilities in adverse conditions—and collectively enrich driving data, thereby enhancing safety and overall intelligent driving performanceThis mutual enhancement from multiple data streams propels the evolution of intelligent driving technologies.

The traffic ecosystem is inherently complex; the integration of AI within this space remains fragmented, addressing only basic functionalities such as traffic signal detection or infraction recognitionTo solve comprehensive challenges, there is a need to weave together disparate elements—from vehicle flow and road conditions to traffic signage—within a higher-dimensional intelligent systemBy driving collaboration between vehicles, roads, and cloud-based models, a cohesive perception of traffic dynamics can emerge, enabling faster, more precise control and decision-making.

Utilizing multi-modal large models can significantly enhance visual analysis precision, generalization abilities, and environmental adaptability

The challenges faced by smaller models can be addressed through the application of these higher-functioning architecturesNevertheless, the real-world application still requires an accompanying set of smaller models that focus on bandwidth efficiency and operational costsAchieving a synergy between large and small models could embody a multi-layered collaborative computation framework, likely serving as a cornerstone for future advancements.

Front-end small models may conduct real-time inspections and rapid analyses, whereas backend large models engage in deeper reasoning and secondary evaluations, thus refining accuracy through human feedback mechanismsThis blended approach leverages the swift capabilities of small models with the predictive strengths of large-scale systems.

Within this road-cloud integration, cloud computing adeptly manages extensive, non-real-time data analysis while supporting business decision-making processes

This facilitation stimulates agile application development across services and optimizes onboard electronic and software systems through comprehensive management and simplification.

Conversely, edge computing is dedicated to real-time data processing, conclusively enhancing vehicle-level intelligent processing and execution capabilitiesEmploying edge small models for quick perception while leveraging cloud-based large models for secondary alerts establishes a robust feedback loop, propelling accurate situational awareness.

In the domain of predicting traffic flow, the standard procedure involves employing time-series data to forecast volume, analyzing a range of intervals—from immediate monitoring to longer-term projectionsHowever, crafting effective traffic forecasts is intricate, conditional upon complex interactions among parameters that can shift due to unpredictable external variables such as weather or accidents

Quality and usability of traffic data need continuous scrutiny.

The road-cloud network envisages that vehicles, roads, and traffic intersections become interconnected intelligent agents, where human directives set objectives while intelligent systems devise stepwise strategiesEach step determines whether to employ a large model or a small model, culminating in a seamless closed-loop operation.

Previously, traffic incident response protocols were often theoretical, lacking in empirical data to illuminate precise causes and severity of issues like congestionSuch blueprints historically served more as repositories of knowledge, with variance in interpretation and execution among different stakeholdersTransitioning to intelligent agent models allows these response plans to evolve into dynamic systems capable of nuanced understanding and responsive engagement, thereby empowering traffic authorities to implement more refined management strategies.

The advancement of intelligent driving technology mirrors a transition from a less sophisticated battlefield to one defined by high-stakes competition involving data and computational resources