LLMs are eroding my software engineering career and I don't know what to do

Perceived erosion of software careers

• Many senior engineers feel core advantages—debugging skill, domain expertise, code quality—are being commoditized as LLMs can now one‑shot bugs, write design docs, and generate working prototypes.
• Some report management pressure to “use more AI” and measure productivity via tokens, lines of code, or ticket counts, making them feel like overseers of “slop farms” rather than builders.
• Others note this pessimism is familiar from earlier waves (offshoring, web, cloud) and argue demand for good engineers remains, just with changed expectations.

Domain expertise vs generalist skills

• One camp: deep domain knowledge (finance, tax, simulation, signal processing, healthcare, regulation) is still a strong moat; LLMs routinely miss subtle rules, edge cases, and institution‑specific practices.
• Opposing view: “domain knowledge is promptable”; broad engineering principles and ability to orchestrate agents will matter more than niche expertise that LLMs can read from docs.
• Several point out domain experts will still be needed to interpret regulation, negotiate with regulators, and own legal risk—tasks not easily handed to models.

Capabilities and limits of LLMs

• Heavy users report genuine gains: faster refactors, complex bug finding, test generation, YAML/config editing, data engineering PoCs, and even nontrivial math/physics code with visual verification.
• At the same time, frontier models still:
  • Hallucinate APIs, regulations, dates, or compliance requirements.
  • Overproduce code, duplicate logic, and ignore architectural patterns unless heavily constrained.
  • Fail badly once pushed outside well‑trod domains or when specs are ambiguous.
• Many emphasize “driver skill”: experienced devs get big wins; juniors and non‑devs often produce fragile, incoherent systems.

Code quality, architecture, and “AI slop”

• Widespread worry that management will accept C‑grade “AI slop” because it’s faster and cheaper, even if technical debt and failure rates quietly rise.
• Others say architecture and clarity matter even more now, because agents and humans both operate better on well‑structured systems, and LLMs can help enforce patterns if guided carefully.
• Several describe a new anti‑pattern: huge PRs one‑shot by agents, followed by long, painful human review cycles that erase most speed gains.

Safety, regulation, and accountability

• In regulated domains (fintech, healthcare, aviation, PCI, GDPR, AML), commenters recount LLMs:
  • Misreading or inventing legal requirements.
  • Proposing non‑compliant designs that would fail audits or create serious liability.
• Consensus here: fully autonomous agentic development is reckless; humans must remain on the hook for code and compliance, and regulators/auditors are not yet AI‑ready.
• Some foresee future “AI slop trials” after serious incidents, which may harden practice around deterministic tools plus human review.

Economic and social implications

• Broad fear that LLMs shrink white‑collar headcount: fewer engineers supporting more product areas, tech salaries normalizing downward, and juniors frozen out.
• Others argue new sectors and tools will appear (as with past tech shifts), and that overall demand for software and automation may rise faster than individual productivity.
• Some extrapolate to “post‑labor” scenarios, questioning what economic system supports billions if most knowledge work is automatable; suggestions range from UBI to wealth taxes to “own capital or land.”

Coping strategies, values, and alternatives

• Common advice:
  • Lean into AI as a “power tool,” not a replacement: learn harnesses, build deterministic tools for agents, focus on architecture, testing, and system design.
  • Shift identity from “coder” to “engineer” or “problem solver”: working on requirements, intent, and constraints rather than keystrokes.
  • Maintain or deepen domain expertise and soft skills (communication, risk judgment, working with legal/compliance, customer understanding).
• A vocal minority is repulsed by the new work style (“vibe coding,” supervising robots) and contemplates exiting to trades, woodworking, farming, or other hands‑on crafts—while others warn those paths are also hard, crowded, and risky.
• Several note a loss of joy: coding becomes less a craft and more a process of editing and supervising machine output, raising concerns about burnout and mass disillusionment before other sectors fully feel AI’s impact.