Large Model R&D Is a Tiny Niche — Aim for Agent Engineering Instead
Large Model R&D is a Small Niche
——Some advice for students job-hunting in the AI era
Author: 7July LogicFrame | CTO Source: WeChat Official Account [林间有风]
A few days ago, a student asked me in the Q&A section about future learning and career direction.
He had self-studied machine learning, image recognition, and natural language processing, researched the Transformer architecture and source code, and even trained a small model. Later, he got into LangChain, LangGraph, RAG, A2A, memory, Skills, and MCP. At his company, he also worked on several real projects: resume conversion, resume retrieval, an AI crawler, and legal regulation retrieval and analysis.
By all accounts, he has learned quite a lot and actually built things, so he should feel more confident. But he still worries about two things: his academic credentials aren't strong enough, so he might not get into an algorithm role; and he's very hesitant about whether to keep diving deep into algorithms, fine-tuning, and LoRA.
This kind of confusion is very common. The more you learn, the more every new buzzword on a job site feels like a death warrant, leaving you anxious and completely ungrounded.
I also frequently take consulting calls from companies, so I know some of the realities on the ground. Today I'll summarize and supplement, sharing with all of you in the hope that it helps.
01 · ATTITUDE
Be thick-skinned when job hunting, carry a kind of carefree nonchalance
The first piece of advice is not technical. It's about attitude. Many students are timid when job hunting, thinking they aren't good enough here or qualified enough there, stammering through interviews and tests with zero confidence.
I think this affects not just one interview outcome, but can have a big impact on your whole life. You have to believe that the world is sometimes really just a makeshift stage. Interviews that ask you to build a nuclear bomb, only to have you tighten screws once you're in — that's the corporate norm.
When I read Wu Xiaobo's Thirty Years of Turbulence over a decade ago, the biggest revelation I got was: if a person wants to succeed, you first have to be "thick-skinned" (I consider this a positive term in certain contexts). Don't care what others say, don't psyche yourself out.
This is an attitude toward life. It's just job hunting — what are you afraid of? Don't be so honest and earnest; learn to package yourself. If you get found out, just laugh it off and move on to the next company.
02 · AUTHORITY
Authority: The AI era has wiped out the vast majority of technical authorities
In the past, in the traditional era, you'd hear someone was from BAT and instantly be in awe — products at a scale of hundreds of billions, high-concurrency systems, distributed systems, all sounded incredibly impressive.
In that era, without practical AI, there were technical barriers between people. But now it's different. Standing behind you is a super Agent that has mastered nearly all the knowledge and technology from decades of internet development, serving you 24/7.
So that technical barrier has been broken. Most technical authorities have vanished. You are on the same level playing field — you just don't believe it yourself.
But you might think some people are still better than you. Yes, someone like Karpathy, for instance. But that superiority is not in technology. In pure technical terms, you can already level the field with the vast majority of people. Those who truly pull ahead of you are stronger in character, cognition, and comprehensive ability — absolutely not in the technology itself.
In the AI era, learning a technical skill is very easy, but continuously improving cognition and literacy remains very difficult. The only way to improve that I can think of, which I've said many times: read, continuously output (can't just input), and pursue practicality.
03 · CAREER
Employment: Large model R&D and Agent development are two different paths
I've found that even after entering the AI era, the nature of some problems hasn't changed at all.
Before, the question I was consulted on most was: Should I study algorithms? Now the question I'm consulted on most is: Do I need to deeply study machine learning, LoRA, Transformer?
This question is both easy and hard to answer. Many people haven't figured out that the current directions are mainly divided into the large model direction and the Agent development direction, and the demand gap between these two paths is huge.
Large model R&D primarily deals with the model itself. You need to study training methods, data quality, distributed training, and model evaluation. Mathematical foundations, paper-reading ability, training experience, and even academic background will all affect your chances of entering this field.
Agent application development deals with a different kind of problem: how to use existing models to solve a real business need within a company. Forget all the flashy Agent buzzwords — Harness, RAG, knowledge graphs, FDE — their purpose is to solve enterprise problems, improve efficiency, and reduce costs.
For example, a company wants to convert tens of thousands of data reports and historical records into a unified format. The difficulty may not be that the model isn't strong enough, but that the original files are wildly diverse, fields have no standards, and the conversion results need to be auditable, traceable, and integrable into existing systems. The model is just one link in the chain. Up front, you have to handle dirty data; downstream, you have to handle permissions, costs, exceptions, human confirmation, and system integration.
Simply put, Agent development requires engineering ability, not mathematical foundations, model training experience, or machine learning.
You have to first clearly distinguish these two different directions.
04 · POSITION
Many people's problem isn't technology, but a lack of clear self-positioning
Last year, I had dinner with a friend who works in a government institution. His unit is excellent, an absolute iron rice bowl. Our conversation turned to the topic of civil service exams. I asked him, what kind of academic backgrounds and schools do the people who get in every year have — 985, 211, PhDs?
He smiled: The largest number of people entering our unit every year are from second-tier and third-tier universities (back in our day, universities were still divided into second-tier and third-tier; they aren't now), roughly schools outside the top 7 in Wuhan, even private colleges.
I was surprised and asked: Why? His reason was simple: These new entrants have extremely strong purpose. From their first year of university, they clearly knew they wanted to take the civil service exam, which direction, which unit — all thought through very clearly. They started preparing in their freshman year and took the national and provincial exams right after graduation.
I thought about it. These people aren't necessarily smarter than others, but they win by recognizing themselves clearly and having a clear self-positioning.
I bring up this example to illustrate that the large model R&D direction is a small niche; it simply doesn't need that many people. The requirements for academic background and credentials are imaginably high. I personally think that without a prestigious university background and a PhD, it's very difficult. If you want to charge toward this direction, you need to weigh yourself carefully.
Many people like to argue: they know so-and-so, with just a bachelor's degree from a non-prestigious university, who is doing large model R&D at some company with an annual salary of whatever. Never mind whether this is even true — even if it is, it's not a common phenomenon. Era, opportunity, luck — multiple uncertain factors intertwined, maybe producing one miracle among hundreds of thousands of people.
This is similar to what I often say: reading and getting an education is the most cost-effective, lowest-cost survival path for ordinary people. There are always people who argue with me: some guy in the village, only finished elementary school, now a big boss earning hundreds of millions a year. By that logic, I should immediately take all my savings and buy lottery tickets. After all, someone always wins tens of millions; I should take a gamble too.
Large model R&D salaries are indeed high, but whether you're suited for it — weigh it yourself. I admire two types of people, and dislike one type:
- One type is talented, brash, and free-spirited;
- Another type recognizes their own average ability, works steadily, and earns honest money;
- But what I dislike most are people of average ability who want a lot, ending up stuck in the middle, complaining about everything.
05 · PRACTICE
Seek truth, seek reality; practice is the primary productive force
The demand for enterprise Agent transformation is enormous. It's not like those messy, pretentious new buzzwords companies throw around. Agentization can almost 100% be confirmed as the future direction — a blue ocean among blue oceans.
So my advice to students is: seek truth, seek reality, with problem-solving as the primary goal. Transformer, supervised learning, fine-tuning, LangChain, LangGraph — put these aside for now, just understand them (you can't be completely ignorant; you must understand them).
Put the bulk of your energy into how to reduce costs and increase efficiency for enterprises, learning through real business scenarios. Let me give two examples.
First problem: CLI-fying CMS functions.
LogicFrame's front-end doesn't have many features, but the back-end is extremely complex, including finance, orders, materials, textbooks, videos, comments, Q&A, notifications, and more. Moreover, this back-end is mostly multi-tenant, with different authors needing to see different views, and permission control is very complex.
Currently, we use the LinCMS system, which is fine. But it's very inconvenient. Because I need to open the CMS, log in, click through menus nested countless layers deep, find a node, upload, and select materials. Efficiency is abysmally low.
So now we are CLI-fying all CMS functions and integrating them into Hermes. If I want to know some data, I just tell Hermes:
Check today's efficiency for me; publish today's course update notes for me; correct the error in chapter xx, section xx for me and republish;
The efficiency improvement is exponential. Hermes is my super personal assistant; I only need to interact with it. This is a very typical Agentization requirement. It's not simple — it requires permission control and multi-level review, because the Agent will make mistakes.
Second problem: Automated operations.
Before, when there were problems, they required manual investigation, manual repair, and manual deployment and operation. We are now gradually trying to use Hermes for operations in the production environment.
Both of these examples are problems we need to solve in our work, and they happen to be useful for improving our understanding of Agents. I have always believed that engineering ability can only be improved through practice. No matter how much machine learning, Transformer, or algorithms you read about, it's useless.
∞ · THE END
Summary
Stop imagining the AI industry as a place only model researchers can enter. The door to this industry is actually very wide, and it's not far from you, but you need to find the right direction: what real, tangible problems can you solve for others, for enterprises — rather than spending all day studying supervised fine-tuning, machine learning, and Transformer.
I have always been a pragmatist. I believe life doesn't have time to delve into things you can't use:
Encounter a problem → Solve the problem → Abstract these problems → Internalize them into your own knowledge
This path is more suitable for most ordinary people.
Life can be idealistic and literary, but for work, let's pursue utility and practicality. That's my advice.
Top 1 from juejin.cn, machine-translated. The original thread is authoritative.
Great analysis.