This week we noticed a "Model Engineering" note on Juejin getting widely bookmarked across the tech community. The author uses a "cooking heat" metaphor to break down AI's three mainstream learning approaches — and the bigger takeaway is this: "How AI learns" is replacing "What AI is" as the new entry-level must-know.

What This Is

The note groups AI learning into three major schools: connectionism (deep learning / contrastive learning), behaviorism (reinforcement learning), and symbolism (knowledge graphs). The three core methods:

  • Supervised learning: learn from labeled examples, like answering exam questions with an answer key
  • Contrastive learning: no labels needed — learn "who resembles whom" (birds of a feather). This is the core method behind large-model pretraining, where massive unlabeled text teaches the model "what the next token should be."
  • Reinforcement learning: trial and error plus feedback, like training a dog

The more valuable move is the author's use of short-video "traffic throttling" to explain PID control (a classic algorithm that auto-adjusts output via proportional, integral, and derivative terms): the proportional term reads current deviation, the integral term tallies historical debt, the derivative term predicts trend. This "skip the formulas, remember the intuition" style is especially friendly to non-technical readers.

Industry View

The pro side: this kind of framework-driven explainer helps product managers and traditional-industry bosses build "model engineering intuition" — so they stop getting snowed by "AI can do anything" marketing.

The counterpoint deserves equal airtime. A senior algorithm engineer in the comments noted that likening reinforcement learning to "training a dog" papers over the brutal difficulty of designing the reward function (the scoring standard that tells the AI what counts as "right") — which happens to be one of the thorniest engineering problems of the large-model era. "Understanding the three schools and actually shipping something are separated by 100 nights of parameter tuning." Moreover, reductive claims like "contrastive learning is the core of large models" can mislead outsiders into overestimating AI's true reasoning capability today.

Impact on Regular People

  • For enterprise IT: when picking an AI vendor, being able to ask "do you use supervised or reinforcement approaches?" gets much closer to substance than just comparing parameter counts.
  • For individual careers: jargon stops being scary. Knowing the intuition (which method fits which scenario) beats memorizing formulas — and lets you actually follow your team's proposals in meetings.
  • For consumer markets: the underlying logic of recommendations and short-video traffic shaping becomes transparent. Users will increasingly sense when they are "being fed by algorithms" and find ways to push back.