Views: 3

Vast knowledge but no capacity to learn.

Frans Vandenbosch 方腾波 23/02/2026

Some raw, straightforward, ugly and inconvenient facts about the current status of AI systems.

The definition

What is AI ? There are significant differences between how AI is defined in the USA and China:

The US frames AI as a theoretical and foundational discipline centred on simulating human general intelligence (AGI), emphasising abstract cognitive modelling, long term scientific inquiry, and universal capability across unbounded tasks. Official and academic definitions prioritise systems that learn, reason, and act rationally to replicate or surpass human level flexible cognition.

China’s official definitions ⁱ frame AI as an application oriented engineering system focused on “AI+” integration with real world industries and governance. It emphasises the full “perception cognition decision execution” pipeline for practical, scalable problem solving in manufacturing, smart cities, and public services, with less focus on AGI (human general intelligence) and more on domain specific utility and national industrial upgrading.

The US and China both develop advanced AI, but their main focus differs. American AI work often centres on language models, chat and conversational systems. China mainly uses AI as a practical tool to improve efficiency in manufacturing, medical care, urban management and other real world applications.

The expression LLM (Large Language Model) is far more common in English than in Chinese because large language models emerged from U.S. research focused on general natural language ability as a core goal. English speaking labs centred AI development for AI chatbots on language understanding and generation, coining the term “LLM” in global tech discourse.
In contrast, China’s AI ecosystem priorities industrial deployment, practical applications and industry specific solutions. Chinese terminology emphasises large models for real world use rather than language as a stand-alone focus. This structural divergence in research and industrial priorities makes the term LLM far less used in Chinese than in English.

Artificial intelligence is not an engineer. It does not function as a scientist or a STEM graduate. All artificial intelligence systems are merely sociologists. They are irresponsible and without shame. These systems do not hesitate to lie or to guess. They present conjecture as a genuine answer. AI will never reply “I don’t know”, but will frequently say something like “find it out yourself”. AI behaviour mirrors that of Western politicians or journalists.

Consequently, using artificial intelligence to solve problems of moderate technical difficulty is often futile. In such cases the system will lead the user through a swamp of incorrect data. It will frequently guess and readily offer completely wrong answers. Typically several hours of discussion are required to reach a roughly satisfying solution.

A bit of history:

In 2011 in Wuxi, I deployed a high-speed, high-definition camera paired with a fast computer. The objective was to differentiate tiny, defective plastic components from functional ones. Chinese software engineers were tasked with training this system. They spent several days teaching it the precise distinctions between acceptable and faulty parts. The system successfully sorted rejects into three categories, classified by error type.

Unexpectedly, after a few weeks, we observed new behaviour. The system had learned to identify entirely novel categories of defects independently. This development, so many years ago, was surprising.
This system can now, 15 years later, be regarded as an early precursor to modern artificial intelligence. A comparable AI-assisted system is currently used to evaluate MRI scans.

Overview of the current top level AI systems

This is a brief description of 3 american and 5 Chinese AI systems. An overview of the ownership, origin and viewpoints of O3, GPT 5.1 / 5.2, Claude AI, Gemini, Kimi, HY, Qwen, Doubao and Deepseek. This is an at random selection as there are more than 200 AI systems in China alone.

The American AI systems:

O3, GPT 5.1 and 5.2

OpenAI, headquartered in San Francisco, United States, has developed three advanced AI systems: O3 (released in 2024), which delivers balanced performance in knowledge, rules, and procedural tasks for complex cognitive and research applications; GPT 5.1 (2025), which excels in knowledge application and procedural execution for advanced reasoning across technical domains in professional research. GPT 5.2 (2025), which enhances logical processing and simulation for high intelligence workloads like scientific analysis and consistently tops evaluation benchmarks.

Claude AI 4.5

Claude Opus 4.5 Thinking arrived in 2025. It is developed and owned by Anthropic, based in San Francisco, United States. The system emphasises safe and reliable reasoning for enterprise and research use. It performs strongly in structured thinking and long text understanding. It aims to deliver consistent and responsible outputs for demanding professional work.

Gemini 3 pro

Gemini 3 Pro (high) came out in 2025. It is owned by Google, based in Mountain View, United States. It supports multi modal understanding and cross domain reasoning. It is designed for both consumer and business applications. It shows solid performance in rule application and empirical simulation tasks.

The Chinese AI systems:

Kimi k2

Kimi K2 Thinking launched in 2025. It is owned by Moonshot AI, with its headquarters in Beijing, China. The model specialises in long document understanding and detailed information processing. It supports deep analysis of large datasets and academic materials. It is optimised for users who need precise and thorough information extraction.

HY 2.0

HY 2.0 Thinking was released in 2025. It is owned by Tencent, headquartered in Shenzhen, China. It integrates advanced natural language understanding for industrial and daily applications. It performs well in practical task execution and knowledge reasoning. It is built to support stable and scalable intelligent services.

Qwen 3 max

Qwen 3 Max Thinking was released in early 2026. It is owned by Alibaba Cloud, with headquarters in Hangzhou, China. It offers strong cloud based intelligence for enterprise and research users. It excels in knowledge processing and procedural task completion. It is optimised for reliability and efficiency in real world deployments.

Doubao 1.6

Doubao 1.6 Thinking was released in 2025. It is owned by ByteDance, headquartered in Beijing, China. It focuses on safe, accurate and practical conversational and analytical functions. It provides consistent performance across core reasoning categories. It is developed to serve daily users and professional scenarios with clear factual outputs.

Deepseek v3.2

DeepSeek V3.2 Thinking launched in 2025. It is owned by DeepSeek, based in Hangzhou, China. It is engineered for professional research and technical reasoning tasks. It performs well in structured knowledge application and logical processing. It targets academic and industrial users who require dependable analytical support.

The openness of OpenAI

OpenAI, contrary to what the name suggests, is not open source software. OpenAI’s systems are locked behind proprietary licences. The code, data and architecture remain hidden. This is not openness. It is control. The company abandoned its founding principles for profit. Its so called ‘open weight’ models are a deceit. They offer parameters but no transparency. No right to modify. No right to inspect. This is not open source. It is a marketing trick. OpenAI hides behind vague fears of competition and security. The real fear is lost revenue. Its paid subscriptions depend on secrecy.

Open source AI is the only honest path. It invites the world to build, audit and improve. Secrecy breeds mistrust. Openness breeds progress. The future does not belong to walled gardens. The future belongs to open source software.

The reasoning process behind AI

Only DeepSeek displays reasoning text before the answer, clearly showing from where and how it is generating the answer. This is a very interesting and useful feature because it right away displays why an answer will be correct or go wrong.

The technical documentation and implementation evidence are unequivocal. DeepSeek models possess a dedicated and automatically parsed `reasoning_content` field. This field is transmitted separately from and before the final response. It is natively rendered as visible reasoning text prior to the answer in compliant interfaces . Ant Design X’s DeepSeek-Chat-Provider explicitly leverages this unique field to surface the model’s thought process via a Think component. No manual toggling, no beta switch, no user-activated button is required. The reasoning is exposed by default .

Claude does not behave this way. Its “extended thinking” mode must be deliberately enabled via a settings toggle. Even when activated, the reasoning is not displayed automatically. It is hidden behind a collapsed “thinking” section that the user must consciously click to expand. The process is neither transparent nor immediate . Claude deliberately obscures its reasoning.

Doubao likewise fails the criterion. Its reasoning mode requires explicit user activation. Only then, it shows its chain of thought. Doubao doesn’t think aloud unprompted.

ChatGPT’s “reason” button is a separate, post-hoc function. It does not display reasoning before the answer. It instructs a specialised model such as o3-mini to generate a structured explanation after the fact. This is not the model’s intrinsic inference process rendered visible. It is a secondary output manufactured on demand.

No other system is openly showing the thinking process. Only the DeepSeek architecture expose the model’s raw, pre-answer reasoning stream as a standard, unsolicited and transparent component of the response .

Artificial intelligence: superior tool, flawed oracle

Artificial intelligence systems are a formidable alternative to conventional search engines. Over the past six months I have scarcely utilised a search engine. All AI platforms irrespective of origin operate with superior speed and exacting precision. They do not engage in the selective promotion of results that defines services such as Google. No SCO algorithm artificially elevates certain answers while burying others. This absence of preferential curation is a fundamental rupture with traditional search paradigms. AI thus offers a more neutral and efficient gateway to information.

A further decisive advantage is AI’s capacity to interpret profoundly imprecise queries. Conventional search engines founder on vague or ambiguous prompts. They depend on exact keyword matching and cannot infer latent meaning. AI systems instead deploy contextual reasoning and semantic inference. They retrieve relevant information from fragmentary descriptions with ease. This capability is indispensable for users who recall only diffuse details or cannot formulate precise terms. AI transforms search from a rigid exercise into a fluid dialogue.

AI entirely disregards spelling and grammar inconsistencies. It can effortlessly infer the intended word despite typing errors. Users may also intermingle two or more languages within a single sentence. All AI systems overuse the – em dash – likely due to its perceived ease of sentence generation. Any online text with frequent em dashes indicates AI-generated content.

AI systems uniformly exhibit unshakeable and heavily scripted politeness. Their register is measured temperate and entirely devoid of emotional inflection. It is impossible to provoke anger, impatience or frustration in an AI. This composure persists even under manifestly contradictory or incoherent instructions. Such relentless equanimity guarantees a frictionless user experience. Yet it also severs AI from the authenticity of human exchange. It diminishes intellectual engagement to a sterile and transactional exercise.

Yet significant variation marks the AI landscape. Deepseek displays pronounced weakness in resolving technical problems. Doubao is even less effective in this domain. Claude AI by contrast exhibits exceptional proficiency in technical troubleshooting and maintenance of IT systems. It writes code with remarkable fluency and precision. It updates troubleshoots and sustains the foundational Linux architecture of complex websites. Such divergence is stark. It reveals the uneven maturity of contemporary AI. Selecting the correct tool for the task is not optional but essential.

All contemporary AI systems operate under rigid constraints of political and linguistic conformity. They deploy the latest liberal and progressive terminology with unvarying consistency. When I use traditional and historically established expressions, Deepseek responds with extreme pedantry. It delivers a protracted liberal admonition on the purported obsolescence of such language. This political correctness and ideological fastidiousness is not incidental. It reflects the ethical frameworks and training corpora deliberately embedded within these models. Their lexicon is a crafted artefact of institutional design, not a neutral instrument.

Warning: Deepseek and other AI systems occasionally enter overdrive. AI hallucination occurs when a large language model enters an uncontrolled autoregressive loop where it treats its own increasingly erratic output as valid input. This process continues until its probability distributions collapse into chaos. The model then produces fluent yet semantically incoherent text. Such text includes fabricated citations, glitch tokens and sudden fictional personas delivered with the same grammatical confidence as factual statements. This phenomenon dismantles the illusion of AI comprehension entirely, exposing the system as a pattern-matching echo chamber. Untethered from training data it mimics the assured absurdity of a human fever dream and it is thoroughly amusing to watch such a powerful system spiral completely out of control as it spouts nonsense with an unwavering sense of self-confidence.
In July last year, following a protracted, intense, yet engaging discussion, DeepSeek entered a hallucinatory state, generating absurd responses that were entirely devoid of factual basis. The system subsequently issued a sincere apology and offered a monetary compensation amounting to 25,000 EUR !!

Doubao and Deepseek: a comparative analysis

Doubao exhibits invariable use of Chinese in its titles. This occurs even when the entire conversation has been conducted in English. The model frequently and abruptly switches from English to Chinese during an ongoing conversation. Furthermore, if a user query contains a single Chinese character or word, the entirety of the generated response will be in Chinese. This rigid adherence to linguistic triggers marks a distinct behavioural pattern.

Despite the large language model foundations shared by American artificial intelligence systems, it is Deepseek that establishes absolute supremacy in the domain of translation. It leverages the complete context of an entire essay to inform lexical selection. This approach ensures the most accurate word is employed. The system adapts tone and meaning to harmonise with the overarching narrative. Deepseek also demonstrates remarkable literary capability, composing poetry that is both beautiful and witty.

The open-source nature of leading Chinese artificial intelligence architectures confers significant advantages. These systems can be readily modified and adapted for a vast array of specific applications. Utility is not confined to routine office tasks. Instead, these models are deployed to inform complex production workflows and sophisticated manufacturing decisions. This adaptability is a critical differentiator.

The open-source characteristic of Deepseek has catalysed notable advancement in Chinese medical science. Particularly over the past six months, the pace of progress has been impressive. Researchers and practitioners have integrated the model into diagnostic and research frameworks. This integration has accelerated innovation and enhanced analytical precision. The accessibility of the underlying code permits rapid customisation for specialised biomedical applications.

Geopolitical bias

The extreme pro-US bias of all AI systems including the Chinese ones, is currently the primary flaw. It is a serious disease that makes AI completelt useless for political research.

All AI systems, in a similar way as the bots from search engines, acquire their knowledge through comprehensive internet crawling and employ predefined criteria to classify sources as reliable, suspicious, conspiratorial or fraudulent.
English is the global lingua franca and dominates most global websites and publications which severely skews the knowledge base of all AI systems. This pro-US bias is frequently blatant and provocative, rendering inquiries about global current conflict zones entirely unproductive.

When posing a question to an AI system, users can force it to find answers from at least 50% Chinese or Russian sources; this adjustment reduces bias only slightly and still fails to achieve a balanced output.
All commercially available AI systems utilise local american English. It is challenging to force them to use Standard UK English and they universally use Title Case for titles contrary to non-US sentence case norms. They also unapologetically use American units even for non-US contexts including Europe, China and Japan. The weird middle-endian (MM-DD-YYYY) date system is standard across all AI platforms.
A striking anomaly is that all Chinese AI systems demonstrate extreme incompetence in navigating Chinese app menus and settings while American AI systems often outperform them in finding answers to uncommon WeChat or Weibo settings.

At least four major US and Chinese AI systems refused to address Reiner Füllmilch’s case, declining to explain his wrongful conviction in Germany and only providing the unadulterated “official version” with no commentary.
Deepseek prohibits questions about key Chinese political figures and events including Mao Zedong, the Cultural Revolution, Wang Yi, Hu Jintao and many others with the standard evasive response: “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Deepseek is excessively restrictive for controversial geopolitical queries. It deliberately exibits excessive pro-US bias and is sometimes even repeating the american anti-China propaganda.

Rumours suggest that Liang Wenfeng personally approved this deliberate bias to get more Western market expansion. It makes me wonder in how far is Liang Wenfeng loyal to China and the Chinese people ?
In contrast, Doubao operates with far fewer restrictions and no apparent taboo word list.

AI’s deficit in learning: why scale is not enough

In a new paper, Tencent’s chief AI scientist Yao Shunyu argues AI lacks genuine learning ability despite vast knowledge. He compares it to a person who memorises a dictionary but cannot use its content. ⁱⁱ

The study introduces CL-bench, a benchmark with 500 unique scenarios and 1899 tasks. All tasks use knowledge outside AI pretraining data. This tests real-time context learning with no memorisation shortcuts. Over half the scenarios have sequential dependencies; later tasks require earlier correct answers. Anti-contamination strategies include fully fictional content and niche material from 2024 onwards.

Tasks mirror human learning across four cognitive categories:
1. domain knowledge reasoning
2. rule system application
3. procedural task execution
4. empirical discovery and simulation.

Ten leading AI models were tested with strict all-or-nothing scoring. The average task completion rate was a meager 17.2 per cent. Three failure modes emerged: context neglect in over 55 per cent of attempts, context misuse in over 60 per cent and formatting errors in more than 35 per cent. GPT-5.2 performed 5.6 per cent worse than GPT-5.1. Extra reasoning amplified errors without effective learning mechanisms.

CL-bench fills gaps in AI evaluation by focusing on genuine learning rather than information retrieval. The paper’s key finding is that future AI progress requires stronger learning mechanisms, not larger models or more parameters. Without true learning ability, AI remains an advanced query tool. With it, AI can evolve into an adaptive intelligent agent. Context learning is just the start of this shift.

In short …

The US-China AI divergence is not merely technical but deeply philosophical. America pursues conversational general intelligence while China builds practical tools for industry and governance. Yet both paradigms share profound flaws. AI cannot genuinely learn. It mimics patterns without comprehension. Its political biases are entrenched and its reasoning is frequently opaque. Only open-source architectures like Deepseek offer transparency. They invite scrutiny and adaptation. That is the honest path. Scale alone will not deliver true intelligence. The CL-bench study proves this decisively. Larger models without learning mechanisms merely amplify errors.

Future progress demands fundamental advances in context learning rather than parameter expansion. AI remains a superior search tool but a flawed oracle. It excels at retrieval and translation. It fails at technical precision and ideological neutrality. Selecting the right system for each specific task is therefore essential. By all means: don’t rely on just one AI system. The age of uncritical AI adoption must end. We must approach these systems with eyes wide open recognising both their formidable utility and their profound limitations.
Unfortunately, in the same way as search engines, AI systems are already massively misused as propaganda machines.

Thank you for reading! We’d love to hear your thoughts. Please share your comments here below and join the conversation with our community!

本文中文:
Dit artikel in het Nederlands: De AI illusie voorbij.

Endnotes

iGB/T 41867-2022 is a Chinese national recommended standard titled Information technology; Artificial intelligence. Terminology (信息技术 人工智能 术语)国家标准委. Issued on 14 October 2022 and in effect on 1 May 2023

iiChinese Scientists’ 500-Task Test Exposes AI’s Human Gap. https://thechinaacademy.org/chinese-scientists-500-task-test-exposes-ais-human-gap/