跪拜 Guibai
← All articles
Frontend · Backend

Large Model R&D Is a Tiny Niche — Aim for Agent Engineering Instead

By 逻辑帧 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The AI job market is flooded with candidates chasing model-research roles that are scarce and credential-gated, while enterprises are desperate for engineers who can integrate existing models into real systems. Misreading the demand means wasting time on ML theory when the paying work is in Agent plumbing, permissions, and dirty-data handling.

Summary

A CTO with hiring experience argues that large-model R&D — training, fine-tuning, Transformer internals — is a small, credential-heavy niche that needs far fewer people than the hype suggests. The bigger, more accessible opportunity is Agent application development: wiring existing models into enterprise workflows to cut costs and boost efficiency. That work demands engineering skill, not math PhDs.

Two in-house examples make the case concrete. CLI-fying a complex multi-tenant CMS so an operator can issue natural-language commands to Hermes replaces endless menu-clicking with a personal assistant that still enforces permissions and multi-level review. Automating production ops through the same agent turns manual investigation and repair into a conversational loop. Both projects teach more about real Agent engineering than any amount of algorithm study.

The advice is blunt: be shameless in interviews, recognize that AI has leveled most technical authority, and stop fetishizing model research. Practical problem-solving, continuous output, and clear self-positioning beat credential anxiety.

Takeaways
Large-model R&D is a small, credential-heavy niche; most people won't get in and don't need to try.
Agent application development is the broad, accessible track — it requires engineering ability, not math or ML training.
AI has wiped out most technical authority; a capable engineer with an AI assistant can now match the output of former elite developers.
Enterprises need Agent integration that handles dirty data, permissions, cost, audit trails, and existing system connections — the model is just one link.
CLI-fying a CMS into a conversational agent (Hermes) replaced deep menu navigation with natural-language commands while keeping multi-tenant permissions and review steps.
Automating production ops through the same agent turned manual investigation and repair into a conversational loop.
Interviewing requires shameless confidence; most companies interview for nuclear physics and hire for screw-tightening.
Practical problem-solving and continuous output build real engineering ability; reading about Transformers and fine-tuning does not.
Conclusions

The framing of large-model R&D as a 'small trade' is a useful corrective to the AI hype cycle, which treats model research as the default career path.

Separating Agent work from model research clarifies a confusion that drives many job seekers to study the wrong things — the industry needs plumbers more than it needs physicists.

The claim that AI has 'wiped out technical authority' is overstated but directionally true: the gap between an average engineer and a former elite one has shrunk dramatically when both have access to the same models.

The civil-service exam analogy is sharp — people with clear, early self-positioning often beat smarter but unfocused competitors, and the same dynamic applies to AI career choices.

The Hermes examples show that real Agent engineering is unglamorous: permissions, multi-level review, dirty data, and system integration, not prompt tweaking.

Concepts & terms
Agent application development
Building software that uses existing large language models to accomplish real business tasks — handling dirty data, permissions, cost, audit trails, and system integration — rather than training or fine-tuning the models themselves.
Large model R&D
The research and engineering of training, fine-tuning, and evaluating large language models directly, requiring deep math, distributed-systems knowledge, and typically advanced academic credentials.
CLI-fying a CMS
Wrapping a graphical content management system's functions behind a command-line or natural-language interface so operators can issue text commands instead of navigating nested menus.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗