Juejin, a Chinese AI community, ran an article last week titled "8 Must-Ask LLM Interview Questions" that pulled in more than 10,000 views; the dominant refrain in the comments was "eliminated on the spot if you can't answer." We've noticed: the technical bar for AI engineers is rising from "can use frameworks" to "understand the underlying math." This shift is also starting to ripple outward to business managers who don't write code.
What This Is
The article is a pass-through guide for AI application-layer engineers, covering Transformer architecture (the core algorithm framework Google proposed in 2017), Self-Attention (the mechanism that lets models understand relationships between tokens), Multi-Head Attention, positional encoding (which lets models understand word order, with RoPE being the current mainstream solution), Tokenization, context windows, sampling strategies, hallucination issues, Function Calling, MCP (Model Context Protocol, the standard protocol for connecting models to external tools), and other foundational modules. It's not issued by a major company; it's a "syllabus" spontaneously compiled by the technical community.
What deserves attention is the syllabus itself: it's a signal that supply and demand for AI engineer roles is shifting fast — enterprise demand for engineers who "understand the underlying layers" has already overtaken demand for those who can "get a demo running."
Industry View
The bullish camp argues: self-attention math isn't that complex, and if you can't follow it, you haven't really built Agents. This bar can filter out people who only know how to call packages, leaving behind engineers who can solve hallucination (models fabricating facts) and long-document processing problems.
But the opposing view deserves more airtime. Several senior engineering leaders have privately said that overemphasizing mathematical derivations will filter out the truly capable hybrid talents who can ship products — what a good AI product lead needs isn't the ability to write formulas, but the ability to break down business problems for the model. Moreover, items on the list like Function Calling and MCP update extremely quickly; they could be obsolete within six months. Codifying them as standardized "must-answer" exam topics carries built-in lag risk.
Impact on Regular People
For Enterprise IT: AI engineer hiring cycles will lengthen and salary premiums will persist; when procuring third-party AI products, more technically savvy evaluators will be needed to gate-keep.
For Individual Careers: Even if you don't code, people who understand the technical narrative will carry more weight in cross-department conversations; those with purely business backgrounds need to find their own irreplaceable niche.
For Consumer Markets: No visible changes in the short term; in the long term, deeper engineering expertise will make AI products more stable, but feature iteration speed may slow.