🌏 中文版
This Week's Behavioral Practice
Story frame: independently killing a two-week-old fine-tuning plan that was headed the wrong way
Austen McDonald, a former Senior Engineering Manager and Hiring Committee Chair at Meta who has run over 1,000 interviews, published an analysis this week (Oct 22) pointing at a shift that's only become visible at the end of 2026: AI-assisted development has devalued "your hands-on experience with a specific technology" as a signal, and interviewers now weigh what Anthropic's own hiring report calls "autonomy and judgment, relative to technical domain expertise and implementation." He also notes that Project Deep Dive rounds — which deep-dive a project you've already shipped, structured like a system-design round except the system is one you built — used to show up only at Staff+ and are now starting to appear at lower levels. The implication for AI engineer behavioral interviews is direct: a story about "how many cross-team meetings I ran" no longer scores well; "I judged it myself, ran the numbers myself, and decided the direction myself" does. Today's story is built around exactly that shift.
Situation: the team needed a new intent-classification model for a support agent. The original plan was to fine-tune an open-source small model, and the PM had already promised the business a two-week turnaround. Ten days in, I found two problems: the training labels were noisy, and even the best checkpoint I could get still lost to a simpler approach — rules plus few-shot classification on the large model we already had — on the edge cases that mattered.
Task: as the engineer owning this classifier, I had to decide whether to grind through the fine-tuning plan to completion or change direction — and changing direction meant admitting the original plan wasn't working before anyone asked.
Action: instead of waiting for the PM to check in on progress, I ran a head-to-head experiment on my own: the current fine-tuned checkpoint versus the rules-plus-few-shot approach, both against the same 200-example validation set. That produced concrete numbers — F1 0.81 for the fine-tuned model, 0.89 for few-shot, with the added benefit that few-shot didn't need retraining every time the data distribution shifted. I booked a 30-minute meeting on my own initiative, led with those numbers, and proposed a compromise: ship few-shot within the original two-week window, and spin up a separate data-cleanup spike for the fine-tuning approach that wouldn't block this launch.
Result: the PM and business stakeholders accepted the pivot, and the launch shipped two days early instead of slipping. The follow-up data-cleanup spike fixed the label-quality problem once and for all, and the team has reused that same "run a head-to-head experiment, then bring numbers into the room" process on similar projects since.
How to tell this story
- Do: make the "on my own initiative" part explicit — you noticed the problem, ran the experiment, and booked the meeting, nobody asked you to. Replace vague claims like "it worked better" with hard numbers (F1 0.81 vs. 0.89, two days early). Late-2026 interviewers — and AI-assisted scoring systems — are especially sensitive to whether a decision was actually yours, which is exactly this week's shift.
- Don't: frame it as "I convinced the team." The point isn't persuasiveness, it's that you did the legwork and quantified the risk before deciding. Don't skip the moment of admitting the original plan wasn't working, either — cutting that out makes it sound like you knew the answer from day one, which undercuts the "judgment under ambiguity" signal interviewers are listening for.
This Week in Review
| Day | Topic | What we practiced | Self-rating |
|---|---|---|---|
| Mon | ML Fundamentals | Separating data-, model-, and evaluation-level fixes for class imbalance; the optimization intuition behind L1 vs. L2 regularization; why reusing validation scores to tune hyperparameters is itself a form of data leakage | {fill in} |
| Tue | Deep Learning & NLP | Self-attention's Q/K/V mechanics and √dₖ scaling; why multi-head attention needs multiple heads; the two different levels at which self-attention's O(n²) cost gets solved (FlashAttention vs. sliding-window attention) | {fill in} |
| Wed | ML System Design | The fixed eight-step framework (problem → metrics → data → features → model → training → serving → monitoring); pairing offline/online metrics with a guardrail; how a feature store fixes training-serving skew | {fill in} |
| Thu | LLM & Agent Engineering | Why agent evaluation has to be layered instead of judged on the final answer alone; RAG's "open-book exam" mental model; why context compaction only pays off above a certain compression ratio | {fill in} |
| Fri | Coding | Reservoir sampling's core invariant (every element ends up with probability k/n); how streaming constraints change which tools you can reach for; why not every ML coding problem should be vectorized on sight | {fill in} |
| Sat | Paper Reading | MAS-Guard: how agentic AI expands the attack surface from text output to executable actions; a two-layer defense of local A-Guard plus a system-level Control Plane; four-tier intervention instead of a binary allow/block | {fill in} |
| Sun | Behavioral & Weekly Review | A STAR story about independently killing a fine-tuning plan headed the wrong way, built around this week's shift toward "autonomy and judgment" in behavioral scoring | {fill in} |
Next Week
The weekly rotation stays the same — ML Fundamentals through Paper Reading, Monday to Saturday, with behavioral and weekly review again on Sunday — but each day's questions and papers will be freshly sourced. All seven topic weights in interview-focus.json are still at the default of 1. If one topic felt noticeably harder this week — say, layering metrics in system design, or dividing responsibilities in multi-agent guardrails — bumping that weight to 2 or 3 will add extra reps on it outside its fixed slot next week. Given this week's read on the shift, it's also worth checking your own behavioral stories for whether they lean on organizational complexity (a pile of cross-team meetings) to signal scope, and trying to recenter them on individual technical judgment instead.
References
- What's changing in behavioral interviews — Austen McDonald, The Behavioral — source for today's framing: "autonomy and judgment over organizational complexity" and "Project Deep Dive moving to lower levels."
- Best AI Tools for Behavioral Interview Prep (2026) — CoPilot Interview — reference for STAR story length and the scoring dimensions (relevance, structure, specificity, impact, ownership, articulation).
- MAS-Guard: Unified Guardrails from Individual Agents to Multi-Agent LLM Systems — arXiv:2610.28397 — source for the Saturday row in this week's recap table.