Click
Follow us by clicking the blue text above
Cover image: Three major engineering paradigms for implementing corporate AI agents: Context Engineering, Agentic Engineering, and Model Engineering
“
𝕀²·ℙarad𝕚g𝕞 Intelligent Square Paradigm· a vibe builder ·Large-scale Paradigm
Why does your company’s AI project always seem to be “misfiring”? This in-depth reflection by Xiaofan, following the series on corporate AI implementation, attempts to provide an answer.From the “museum” dilemma of RAG, to the tragedy of the “legless brain” of the Agent, and then to the truth of Palantir’s “clumsy efforts,” layer by layer, it points to the core:The true wisdom of the enterprise is not in the system, but in the human brain.The ultimate remedy proposed in the article is not a new tool, but a new paradigm—initiating corporate knowledge production and letting AI be the “apprentice” for long-term learning; this is the only way out.
Friends, the vibe in the AI circle has indeed been a bit off lately.
Last year, we were still talking about the “dawn of AGI.” This year, VCs are sharing the “Diamond Sutra” and analyses of real estate by Ou Shen.
The reason is simple: too many AI projects that were once highly anticipated have died quietly and ungracefully. I recently attended such a “funeral.” The deceased was our familiar old friend—Mr. Li, the CAIO, whose once-glorious AI empowerment project was said to “lead the next decade.”
At the farewell dinner that day, a PhD from the project team, once a “proud child of heaven” with shining eyes, now looked vacant, downing drinks and recounting how this tens of millions of dollars “accident” gradually led to its destined, sorrowful end.
The First Funeral: RAG Project—Died from “Trying to Recreate a Real War in a Museum”
“At first, we thought RAG was an easy question,” the PhD said with a bitter smile, “Isn’t it just about shredding documents and feeding them to the dog… oh no, feeding them to AI?”
“We scanned the company’s 20-year ‘official history’—product manuals, SOPs, market reports—into the AI’s brain. We thought we had built an all-knowing ‘digital historian.’”
“As a result, when sales asked it: ‘What’s the key to securing Client A?’ it dug up a three-year-old market report saying ‘Client A is price-sensitive.’ This answer is more outdated than my grandmother’s love story.”
“Because it had no idea that Sales Champion Old Wang had just signed the contract for next quarter while singing a duet of ‘Brotherhood’ with Client A’s procurement at KTV. The key to signing the deal wasn’t lowering the price; it was that off-key rendition of ‘Brotherhood.’”
Do you see it clearly? The reason the RAG project was the first to fall was that what we fed it was all beautified, filtered, and severely outdated “specimen knowledge.” We tried to use an AI that had only read history textbooks to fight a rapidly changing modern war.
This is not called AI; this is called “AI archaeology.”
For the detailed accident claims report, see: Burst the corporate AI Agent bubble; it only takes a real RAG project—first clarify the corporate knowledge base documents before proceeding…
The Second Funeral: Agent Project—Died from “Giving a Legless Brain Orders to Run”
“Climbing out of the RAG pit, we thought we needed to make AI ‘move.’ So, we launched the Agent,” the PhD’s eyes dimmed further, “Then we found that we had invited a ‘super brain’ but forgot to equip it with hands, feet, eyes, and ears.”
“We wanted the Agent to automatically approve a ‘business trip application.’ It read the SOP and knew to look for the ‘department head.’ But it didn’t know that the head’s WeChat avatar was a lotus flower, indicating he was recently into string beads and in a bad mood, so it was best not to bother him.”
“It wanted to call the OA system but found that the interface left by the outsourced team ten years ago was a ‘urban legend’—no one could find the documentation.”
This is the dilemma of the Agent. We took a “lion” (public domain LLM) that grew up on the vast grasslands of the “public internet” and forcibly airdropped it into the “corporate intranet,” a tropical rainforest filled with traps, walls, and indigenous tribes.
No matter how powerful it is, it cannot perform.We are not doing AI; we are doing “cross-world travel,” and then expecting the protagonist to automatically adapt to all the rules of the new environment.
For the detailed accident claims report, see: Just climbed out of the RAG pit, and then completely lost in the Agent’s “triple door”—the real experience of AI engineers crossing the tribulation
The Third Funeral (and the most core cause of death): Palantir’s “Imitation Show”—Died from “Trying to Defeat Magic with Magic”
“In the end, we were forced into a dead end,” the PhD downed the last sip of his drink, his voice trembling slightly, “Mr. Li said we were all wrong before; we need to learn from Palantir, we need to do ‘Ontology,’ we need to create truly powerful corporate AI!” (Note:Palantir Technologies Inc. is a US software and services company headquartered in Denver, Colorado, known for its technology in the field of big data analysis. The company primarily provides solutions and services to government agencies and financial institutions.)
“We thought we had found that lost ‘magic book.’”
“And then?” I asked.
“Then we found,” the PhD said slowly, “Palantir’s underwear has no magic; it’s just a bunch of ‘human translation machines’ that are smarter than us!”
“We thought Palantir’s success was the success of AI. Later we realized,that was fundamentally the victory of ‘consulting,’ the victory of ‘humans’! The first 20 years of deep cultivation in the industry is the soil in which today’s LLM can take root.”
“Its Ontology was not ‘automatically’ generated by AI. It was its top consultants, called FDE, who, like ‘Qin Shi Huang,’ delved into every business tribe of the client, spending years to forcefully conduct a ‘unification of language and standards’ through half-threats and half-compromises!”
“They used ‘human’ wisdom to unify the sales department’s ‘deck’ and the marketing department’s ‘presentation materials’ into something that machines can understand asObject: Presentation.”“They used ‘human’ consensus to map the boss’s instruction of ‘go make a PPT’ and the employee’s statement of ‘I’ll write a PPT’ intoAction: Perform_Edit().”
“They first used countless ‘AI laborers’ to painstakingly repair a ‘track’ that connects heaven and earth, and only then let the ‘train’ of AI run on it.”
“And what about us?” the PhD pointed at himself, “We are still in a swamp for the foundation, yet we want the train to take off directly. We tried to use a fictional ‘magic’ (the fantasy of Palantir) to defeat the ‘magic’ in front of us (the reality of not being able to manage corporate AI).”
As a result, we were knocked down by reality with a punch.
The accident site claims report and the only way out: abandon fantasy and start farming.
So, what is the final “autopsy report” of these funerals?
Cause of death:Strategically exerting effort in the wrong direction.
All of us made a fatal mistake:We always want to have a “smart AI” first and then expect it to help us sort out the “chaotic business.”
We imagine an aircraft carrier in the ocean (LLM), but lack the “fighter jet” (a digital engineer with high cognitive density who understands business abstractions) that can take off and land on the deck.We have the best tractor drivers and the best car mechanics. But we lack that “cross-border genius” who understands both tractors and aircraft carrier takeoffs and can abstract the experience of driving a tractor into the skills of taking off from an aircraft carrier.
And the success of Palantir, and our only way out, is precisely the opposite:We must first have a mechanism that continuously transforms “chaotic business” into “knowledge that AI can understand” before we can have a truly “smart” BI by AI.
This mechanism is the Corporate Knowledge Production Agent.
This is no longer the method of “invoking the gods.” This is the method of “dispatching countless digital ‘anthropologists’ and ‘war correspondents’ to our ‘business tribes’ for long-term field research.”
-
Abandon the one-step “aircraft carrier dream” and admit that we only have “tractors.” But our goal is not to use tractors to pull aircraft carriers, but to use tractors to cultivate this wasteland inch by inch, creating an “AI train” on land before thinking about conquering the ocean.
-
Let AI be an “apprentice” first, not an “expert.” Its first task is not to “predict the future,” but to “understand the present.” Let it eavesdrop on Old Wang’s phone calls 24/7, let it peek at Old Li’s WeChat, and convert the smoky “wisdom from the wine table” into structured “field logs.”
-
First let knowledge “come alive,” then talk about intelligence “running.” When these “field logs” are numerous enough, and when AI’s brain begins to “emerge” its understanding of this land, then the “AI train” we dream of will have its first track to run on, even if it’s just a muddy one.
This is slow, clumsy, and not sexy. It requires the patience of daily progress rather than the fantasy of instant success.
But friends, please believe me. After burying so many “genius” AI projects, you will find that the slowest, clumsiest, and most down-to-earth path is often the only way to survive.
#Large-scale Paradigm An AI entrepreneur, a vibe builder, caffeine in the blood, code at the fingertips, anxiety on the forehead
#Artificial Intelligence #Intelligent Agents #Agentic LLM #BI by AI #Palantir #Agentic Engineering #Context Engineering #Context Engineering #Agent Engineering #Corporate Intelligence #Business Intelligence #BI #Corporate AI Apprenticeship #Corporate Knowledge Production
by Large-scale Paradigm, 𝕀²·ℙarad𝕚g𝕞Another way of writing
Scan to join the group, connect with friends!
AI Square Paradigm Think Tank·Large-scale Paradigm