2027 LLM Campus Hiring: Base Roles Hit 3M RMB, Application Layer Commoditizes

Annual Snapshot: Fifth Edition of Campus Recruitment Analysis

The author has published a campus-recruitment analysis every autumn since 2023. This fifth installment treats each article as an annual snapshot of the large-model hiring market, making visible which segments are rising and which are shutting down.

Salary Update: Base-Layer Packages Surge to 3M+ RMB

Last year the author noted that base-model (基座) fresh-graduate packages started at 1.5M RMB. For 2027 the starting figure has doubled to 3M RMB , and in competitive situations — multiple top teams chasing the same PhD with a representative work — the number can be pushed higher on the spot. Offers now include signing bonuses, retention bonuses, and non-compete compensation, terms previously seen only in finance and sports. The author cites a quote attributed to "Liang Sheng": "I'm not afraid of anything, I'm just afraid you'll poach my people."

Why can they pay? Base-model companies' funding and revenue are deeply bound to the large-model narrative; model capability is their stock price. Hiring talent is effectively buying back their own equity and options.

Market Bifurcation: A Cliff, Not a Slope

The market has split into two distinct layers with no middle ground:

Base layer (基座层) : few hires, extremely high bar, very high pay. Campus quotas are countable on one hand and often filled by internal interns.

Application layer (应用层) : many hires, low bar, low pay. Job boards show pages of "large model application" roles requiring RAG and Agent framework familiarity, yet salaries match ordinary algorithm or search/ad/recommendation roles from three years ago.

The previously fuzzy middle — industry-specific models, private-deployment models, vertical models — has been squeezed out. Teams that pursued those directions have largely disbanded.

Reproductive Isolation: The Base-Layer Door Is Nearly Shut

Borrowing the biological term "reproductive isolation," the author describes base-layer hiring as "inbreeding within the circle, sterilization toward outsiders." Without a base-model internship, a paper with real impact, or an open-source project with significant stars , a resume rarely passes the initial screen; even internal referrals fail because the talent pool is too deep.

Anecdote: a top-university master's graduate with papers in search/ad/recommendation applied broadly to base roles and received zero responses. The author's blunt advice: "Last year was the best time to enter; the year before was second best."

Reason: training a model costs tens of millions of RMB. No one will hand that budget to a "theoretically high-potential" stranger. Potential is worthless against a ten-million-dollar experiment. The only credible signals are a base internship (proving you've been vetted by the circle) or a high-impact work (proving you can shine on your own).

The sole remaining crack: intern-to-full-time conversion . This is the last ticket for ordinary candidates.

Agent Development Has Become an Engineering Discipline

Last year Agent work was framed as algorithm research (ReAct, planning). This year the author observes that Agent development is gradually becoming a developer's job . A well-known Agent startup (which later failed due to overseas financing issues) hired only developers, not traditional algorithm engineers, and their technical judgment was a year ahead.

The hard problems in Agent deployment are not decision logic (the model handles that) but engineering challenges : context management, tool-call fallbacks, state-machine design, long-task recovery, cost control, monitoring/alerting, sandbox operations. These match the skill set of engineers who built business systems and middleware. Pure algorithm backgrounds have no advantage here — tuning loss matters less than writing solid retries and idempotency.

Consequently, Java and Go developers are entering the large-model field en masse. "Agent engineer" titles represent the widest stream of developer roles this cycle. For developers, this is a real opportunity: ten years of engineering intuition becomes core competitiveness; the ticket hasn't expired, only the gate has moved.

Role Boundaries Dissolve: Infra = Algorithm = Data = Infra

The author highlights a converging consensus: "infra is algorithm, algorithm is data, data is infra." Last year base teams had clear silos: infra (training frameworks, parallelism, memory optimization), algorithm (architecture, loss, evaluation), data (cleaning pipelines). This year those lines have vanished.

Infra needs algorithm knowledge : improving MFU or MoE load balancing requires understanding routing and training dynamics; optimizing throughput without algorithm insight is useless. Base infra interviews now ask algorithm questions, prompting jokes that they're borrowing the neighbor team's questions.

Algorithm has become data engineering : pre-training bottlenecks are no longer architectural — architecture innovation is saturated. Winners are determined by data mix ratios, deduplication quality, and pipeline robustness. Several leaderboard-topping models this year made almost zero architectural changes; gains came entirely from data engineering. Algorithm engineers now spend most time writing cleaning scripts; the "algorithm" label is increasingly nominal.

Data requires infra expertise : corpora reach tens of trillions of tokens. A single deduplication strategy change takes three days on the cluster. Without distributed systems, scheduling, and storage I/O knowledge, a cleaning script cannot even touch the full dataset, producing local optima unknowingly.

Implication: stop preparing for interviews by role name. Infra candidates must study algorithm; algorithm candidates must study data; data candidates must study systems. The market is hiring one hybrid profile .

Application-Layer Algorithm Work Degrades to Prompt Engineering

Last year the author claimed application-layer algorithm work was essentially prompt writing, which drew pushback citing RAG, fine-tuning, and workflow orchestration. This year those differentiators have disappeared: RAG is absorbed by long-context and internalized retrieval; fine-tuning is wrapped as vendor APIs; workflow orchestration is visualised into drag-and-drop platforms that ops staff can deploy. The daily loop for algorithm hires is now bad-case analysis → prompt tweak → eval run → repeat .

"You write three lines of prompt with Claude Code anyway." An application-team friend described his role precisely as "prompt engineer" despite the title "LLM Algorithm Expert," delivering the line with a blank expression.

Prompt engineering is important — a good prompt can swing metrics by 10+ points — but importance ≠ scarcity. The supply of people who can write prompts stretches "from Guomao to Xierqi" (Beijing tech hubs). Important but non-scarce work commands no price premium because replacement is trivial.

The three traits — many hires, low bar, low pay — form a self-reinforcing structure: low bar → more applicants → intensified competition → weaker bargaining power. Interview question banks have turned into rote memorization (Transformer attention formula, RoPE rotation matrix derivation); three years ago these filtered half the candidates, now everyone scores full marks. Screening fails, so pricing fails.

Actionable Advice by Candidate Profile

Early-stage students (freshman/sophomore, 1-2 years window)

If you want base-layer: secure a base-model internship at all costs . Ignore pay, city, or whether it's grunt work. Get inside. The door crack is only the internship path. Three months of grunt work inside beats ten toy models built outside.

Imminent graduates without internship or representative work

Do not bash your head against the base layer. You'll waste your time window. Application roles are not shameful; get on the vehicle first, switch seats later. Application roles remain the largest-capacity, most background-tolerant large-model positions — even non-elite undergraduates with serious prep have a shot, which is unthinkable on the base side. Once inside, stay sensitive to model-level shifts and wait for the next reshuffle (the industry reshuffles every ~3 years; you'll see 2-3 more in your career).

Developer background

Agent engineering roles are tailor-made for you. This is likely the highest-ROI ticket for ordinary backgrounds in this cycle. Don't hesitate.

Universal trap warning

When a JD reads "LLM Application Algorithm Engineer," ask exactly what the daily work is . If the answer is prompt tuning and eval runs, reconsider how that experience will look on your resume in two years. Titles lie; resumes don't.

Era Shift: From Dividend to Reshuffle to Inheritance

2023: "dividend era" — everyone shouted upside; you just had to join the queue. 2025: "reshuffle era" — half the small base-model companies died; talent consolidated into the top three. 2027: "inheritance era" — the dividend period has ended; the inheritance period has begun. Tickets are no longer issued; they become hereditary.

Personal effort matters, but so does historical timing. Whether you caught the 2023 train affects outcomes more than intelligence or diligence. Recognizing this isn't pessimism — it tells you where to apply force: on the one door crack still open.

Windows never give advance notice of closing. Everyone realizes only after they've shut.

Summary

Base-layer door nearly closed; only internship crack remains. Those with a window must fight to enter.

Can't get in? Take an application role first; don't be picky about the seat.

Agent engineering is the developer's best ticket this cycle.

Algorithm engineers: don't treat prompt writing as the destination.

Job titles lie; resumes tell the truth.