This article compares technical tutorials, product and documentation pages, research papers, and institutional articles recorded, published, or listed between July 22 and September 6, 2026. The AWS tutorial introduces an assistant that accepts orders through text, voice messages, and live voice calls. The three channels share customer memory, and the channel layer is separated from the ordering layer.[S5] A Baidu AI Cloud page describes Qianfan as a one-stop enterprise-grade foundation model service platform centered on Agents, and lists AI Workflow orchestration, Agent application deployment, and multiple model entry points on the same page; their appearance together does not mean that technical integration has been completed.[S6] Baichuan's documentation page lists sections for medical foundation models, general-purpose foundation models, knowledge bases, embedding models, file management, and MCP, but the page does not provide enough information to confirm the specific functionality or availability of each section.[S7] The abstract of the “Compile by Training” paper introduces a method for converting natural-language task specifications into local neural functions that can be saved, versioned, and composed. The authors report 83.6% semantic accuracy on FuzzyBench-Hard, with compilation taking approximately one minute.[S4] The abstract does not provide sufficient detail to independently assess the experimental design, statistical stability, or full costs. These materials are suitable for comparing specific designs and source-reported metrics, but they cannot establish the overall direction of competition in global and Chinese markets, nor can public articles, APIs, or documentation sections be treated as equivalent to an open-source ecosystem.
The Comparison Focuses Not Just on Models, but Also on How Tasks Run
This comparison uses five criteria: whether models are called in isolation or appear in applications with memory or business processes; whether repetitive tasks can be converted into neural functions for local reuse; whether the channel layer and business layer are separate; whether public releases specify weights, licenses, and migration requirements; and whether pages present concrete mechanisms or product and documentation entry points. The first three criteria compare the operating approaches explicitly demonstrated in the materials. The last two examine openness and portability; conditions that sources do not specify cannot be assumed to have been met.
These criteria help explain architectural differences visible in the materials, rather than determine whether enterprise purchasing has broadly shifted toward a particular model. The AWS tutorial explains how a cross-channel ordering assistant sets up a channel layer, an ordering layer, and shared memory.[S5] The abstract of the Compile by Training paper states that the generated functions can run independently of the teacher model at runtime, but this result does not directly represent the overall costs of different enterprise tasks.[S4] The Baidu page lists its positioning as an Agent platform, AI Workflow orchestration, Agent application deployment, and model entry points on the same page; Baichuan's documentation lists sections for models, knowledge bases, file management, and MCP.[S6][S7] These pages do not provide utilization rates, long-term reliability, or cross-platform migration costs. The brief text on the Baichuan page is also insufficient to confirm the specific functionality and availability of each section.[S7]
- Architecture comparison: Examine model calls, state persistence, tool access, and business processes separately, without assuming that related entry points have been technically integrated simply because they appear together.
- Cost comparison: Determining which approach suits a specific task requires accounting for compilation, remote calls, hardware, maintenance, retraining, and failure handling.
- Openness comparison: Public articles, APIs, or documentation entry points cannot replace an examination of weights, licenses, export formats, and requirements for independent operation.
- Reliability comparison: A tutorial's demonstration of a feature does not mean that accuracy, state consistency, and error recovery have been validated in long-term operation.
What Conclusions Can Different Sources Support?
Paper abstracts, technical tutorials, product pages, documentation sections, and case lists provide different levels of detail. The following table lists verifiable content and its scope by source to avoid inferring broad industry trends from individual cases.
| Comparison Subject | Verifiable Facts | Supported Conclusions | Limits of Applicability |
|---|---|---|---|
| Remote Agent application | The AWS tutorial introduces a WhatsApp ordering assistant based on Amazon Bedrock AgentCore and Amazon Nova 2. It accepts text, voice messages, and live voice calls, separates the channel layer from the ordering layer, and uses shared memory to recognize the same customer across all three channels.[S5] | In this case, the three interaction channels share customer memory, while channel processing and the ordering business logic are organized into separate layers.[S5] | This is a tutorial for a specific application. The materials do not provide development hours, operating costs, long-term accuracy, or failure data, so they cannot confirm whether this layering reduces development or migration work.[S5] |
| Local neural functions | During compilation, Compile by Training uses a teacher model to generate task examples, then trains small adapters for a compact interpreter. The paper abstract states that the generated functions no longer use the teacher model at runtime and can be saved, versioned, and composed.[S4] | The authors introduce and demonstrate a method for converting natural-language task specifications into reusable local neural functions.[S4] | The authors report that the method achieves 83.6% semantic accuracy on FuzzyBench-Hard, which the paper describes as the subset on which the Program-as-Weights fast compiler produced no exact matches. Compile by Training takes approximately one minute, compared with seconds for the fast compiler.[S4] This information comes from the paper abstract and does not allow independent verification of the experimental design or statistical stability, nor does it represent other tasks or production environments. |
| Cloud Agent entry points in China | The Baidu AI Cloud page describes Qianfan as a one-stop enterprise-grade foundation model service platform centered on Agents, and lists AI Workflow orchestration, Hermes Agent application deployment, and multiple model entry points.[S6] | The page brings together its Agent positioning and entry points for workflows, application deployment, and model selection.[S6] | July 22, 2026, is the date the evidence entry was recorded and cannot be directly treated as the page's publication date. The appearance of these entry points together also does not, on its own, establish technical integration, customer scale, reliability, end-to-end costs, or migration difficulty.[S6] |
| Multimodal and industry cases | Hugging Face published an article titled “NeoMME: an efficient Multimodal-native and Multilingual Encoder.”[S2] Titles of cases related to assistive robotics at the University of Pittsburgh and Genesis Mission research projects appeared in a list on Meta's website.[S8][S9] | These public titles concern multimodal and multilingual encoding, assistive robotics, and research projects.[S2][S8][S9] | The titles do not establish model performance, deployment scale, project outcomes, weight availability, or licensing terms.[S2][S8][S9] |
| Model and tool documentation in China | Baichuan's documentation page lists sections for medical foundation models, general-purpose foundation models, knowledge bases, embedding models, file management, and MCP.[S7] | The direct conclusion the page supports is that these section names appear in the documentation navigation.[S7] | September 6, 2026, is the recording date marked on the evidence entry and should not be treated directly as the page's publication date. The brief page text does not confirm the specific functionality, availability, performance, adoption, integration costs, or cross-platform migration capabilities of each section.[S7] |
The Components of Task Systems Are Visible, but Do Not Establish an Industry-Wide Shift
The model in the AWS tutorial is not presented in isolation: the application handles text, voice messages, and live calls, separates the channel layer from the ordering layer, and uses shared memory to recognize customers across channels.[S5] This confirms that the case includes components such as channels, memory, and ordering business logic. However, the source provides no user-experience metrics, so it does not establish how much this combination improves accuracy, efficiency, or customer experience.
The Baidu page presents a product platform. It describes Qianfan as a one-stop enterprise-grade foundation model service platform centered on Agents and lists AI Workflow, Agent application deployment, and multiple model entry points on the same page.[S6] This shows that the relevant positioning and entry points are presented together, but the page information alone cannot determine whether they already form a technically complete, commercially unified infrastructure system.
If an application's memory formats, tool permissions, and workflow definitions can only be used within a single platform, cross-platform migration may require converting state or rewriting processes. If these elements support export in common formats, less adaptation may be needed. The current materials contain no cross-platform migration tests, so whether the control layer constitutes a significant constraint needs to be tested through export capabilities, rewriting hours, and post-migration failure rates.
These cases also do not prove that enterprise purchasing has broadly shifted from individual models to complete task systems. Testing for such a change requires procurement contract structures, adoption rates, spending composition, or market data over time. At present, the evidence confirms only that the AWS tutorial demonstrates an application design with shared memory and layered business logic,[S5] while the Baidu page brings together entry points for Agents, workflows, application deployment, and models.[S6]
Cloud Orchestration and Local Reuse Change Cost Components, but Do Not Establish Cost Savings
The AWS case demonstrates a design in which three interaction channels share customer memory and the channel layer is separated from the ordering layer.[S5] Whether this reduces integration and maintenance work depends on interface consistency, development practices, state management, and subsequent failure data; the tutorial's structural description alone cannot answer that question.
Compile by Training proposes another operating approach: a teacher model generates task examples during compilation, after which small adapters are trained for a compact interpreter. The paper abstract states that the generated functions no longer use the teacher model at runtime and can be saved, versioned, and composed like ordinary software.[S4] The authors report that the method achieves 83.6% semantic accuracy on FuzzyBench-Hard, with compilation taking approximately one minute. The Program-as-Weights fast compiler takes seconds but produced no exact matches on this subset.[S4]
The paper abstract reports and demonstrates the generation of functions that do not use the teacher model at runtime in the experiments described, but it does not provide enough detail to independently verify the experimental design and statistical stability, nor does it provide a full cost comparison covering hardware, maintenance, retraining, and failure handling.[S4] Lower unit costs can therefore be treated only as a conditional proposition that needs testing: compiling first and then reusing may reduce unit costs only if tasks are called frequently, input distributions are relatively stable, accuracy meets business requirements, and local operation and maintenance cost less than continued remote calls. If tasks change frequently, require repeated retraining, or entail high error-handling costs, the result may be the opposite.
The choice between cloud orchestration and local functions also needs to be tested for each scenario. If a task depends on continuously updated models, external tools, and shared data, the update costs and call overhead of remote solutions can be compared. If network latency, data control, or service dependencies are the main constraints, the hardware costs, accuracy, and maintenance burden of local functions should be tested. The current materials do not complete such a comparison and cannot determine which approach is more flexible or economical overall.
Differences in Source Formats Do Not Directly Explain Industry Differences Between China and Other Countries
The AWS source is a technical tutorial, so architectural descriptions such as channel separation and shared memory can be verified from it.[S5] The Compile by Training source is a paper abstract that presents the method, 83.6% semantic accuracy on FuzzyBench-Hard, and an approximately one-minute compilation time, but it lacks sufficient detail for independent assessment of the experimental design and statistical stability.[S4] An MIT institutional article describes collaboration between MIT-affiliated individuals and the MIT-IBM Computing Research Lab, stating that its goal is to bring rigorous theory into production systems.[S3]
The Baidu source is a product and event page. What can be verified is Qianfan's positioning as an Agent platform and the appearance of AI Workflow, Agent application deployment, and a model catalog on the page.[S6] The Baichuan source is a documentation page from which only section names—medical foundation models, general-purpose foundation models, knowledge bases, embedding models, file management, and MCP—can be directly verified; the sections' functionality or availability cannot be further confirmed.[S7] These two types of pages provide information at a different level from technical tutorials and paper abstracts.
It is therefore not valid to generalize that international companies focus on mechanisms while Chinese companies focus on integration simply because the AWS materials are more specific and the Baidu materials list more entry points. Differences in source type and purpose may affect the level of detail in public information, but the existing materials do not prove that this is what causes the differences. Without consistently defined data on costs, failure rates, call volumes, and cross-cloud migration tests, these materials also cannot be used to compare the maturity or efficiency of industries in China and elsewhere.
The Baidu page lists models including ERNIE 5.1, DeepSeek-V4, and ERNIE 5.0, accompanied by vendor descriptions of pretraining costs, context efficiency, coding, and Agent capabilities.[S6] This content shows only which dimensions the page uses to describe products. Because the materials provide no uniform testing methodology, they cannot be used to establish a comparative performance ranking.
Public Communication, Interface Entry Points, and Open-Source Ecosystems Are Not the Same Thing
The NeoMME article on Hugging Face confirms that the relevant title was published through a public platform.[S2] The Meta list confirms that titles of cases related to assistive robotics and research projects appeared on its website.[S8][S9] These titles do not specify whether model weights can be downloaded, what uses the licenses permit, or whether the relevant applications can run independently of the original provider's services.
Baichuan's documentation lists section names for models, knowledge bases, embeddings, file management, and MCP,[S7] while the Baidu page lists entry points for models, AI Workflow, and Agent application deployment.[S6] The number of sections or entry points does not directly indicate integration difficulty, nor does it establish specific functionality, availability, exit costs, or migration costs.
Assessing openness requires examining weight availability, licenses, export formats, supported hardware, third-party replication, and conditions for cross-platform operation separately. The current entries support only the existence of article titles, case titles, and documentation sections; they cannot be used to produce a uniform openness ranking of global or Chinese ecosystems.
If the evaluation goal is to reduce self-deployment and maintenance work, the service scope, fees, and dependency conditions of hosted interfaces need to be compared. If the goal is to gain greater control over deployment, weights, licenses, hardware compatibility, and operational responsibilities need to be examined. The current materials do not provide this comparative data, so the suitability of an approach cannot be judged solely by whether articles are published, interfaces are offered, or documentation sections are provided.
Three Falsifiable Propositions Should Be Tested Next
First, if the Agent layer becomes a significant migration constraint, additional costs such as workflow rewriting, memory conversion, and permission-system redesign should be observable when migrating across models or platforms. Future work can record export formats, rewriting hours, and post-migration failure rates; if this state can be migrated at low cost, the proposition does not hold.
Second, if high-frequency, stable tasks are suitable for conversion into local neural functions, Compile by Training should achieve the required accuracy on target business data and show lower unit costs as call volume increases after accounting for compilation, hardware, updates, monitoring, retraining, and failure handling. The paper reports only 83.6% semantic accuracy on FuzzyBench-Hard and an approximately one-minute compilation time; it does not validate this full-cost proposition.[S4] If task changes lead to frequent retraining, or if this accuracy does not meet specific business requirements, the range of applications for this approach will narrow.
Third, if open ecosystems are assessed in terms of portability, future evaluations should examine weights, licenses, export formats, supported hardware, third-party replication, and cross-cloud migration costs together. Public articles, case titles, or documentation sections alone do not establish that an application can operate independently of the original provider's services.[S2][S7][S8][S9]
What can currently be confirmed is a set of specific presentations and implementations: the AWS tutorial demonstrates cross-channel shared memory and layered business logic,[S5] the Baidu page brings together its positioning as an Agent platform, AI Workflow, and application deployment entry points,[S6] Baichuan's documentation lists section names including MCP, models, and knowledge bases,[S7] and the abstract of the Compile by Training paper reports and demonstrates local neural functions that do not use the teacher model at runtime.[S4] This information cannot determine competitive outcomes in global and Chinese markets, nor can it support conclusions about revenue, market share, return on investment, or overall cost savings.
Keep following engineering practices
Subscribe to new articles with an RSS reader, or share this article with someone working on similar problems.
Subscribe to RSS updates