Quartz 5

Goal-Oriented Research is Its Own Discipline

Properties1
tagsagent, llm

3 min read

Summary

A two-year practitioner’s reflection on why startup research is not a compromise between academic research and engineering but a distinct discipline. The author proposes a dual-timescale framing: the long-term research vision is framed as uncertain and resourced as a research bet, while each sprint is framed as a deterministic engineering problem with defined sub-goals and specific people/time commitments. This framing changes allocation, communication, and the signal/noise distinction between sprint failures and vision failures.

基於兩年 MTS 經驗的反思：創業研究不是學術研究與工程的折衷，而是獨立實踐。作者提出雙時間尺度框架：長期研究願景（不確定路徑）作為研究賭注資源化；每個 sprint 作為工程問題處理（明確子目標、具體人員/時間承諾）。此框架改變了資源分配、溝通語域與訊噪判斷。

Key Points

  • The engineering-academic “spectrum” framing is wrong — startup research is not a tradeoff on the same axis
  • Long-term goal: frame as research vision (uncertain path, committed budget against uncertain payoff)
  • Current sprint: frame as engineering problem (well-defined sub-problem, specific people, specific time, measurable whether sub-problem was solved)
  • Two-register communication rule: vision language for customers/executives; sprint language for executing team — confusing them is a common failure mode
  • Pivot signal: single sprint failure = information for next sprint; multiple similar sprint failures = signal to adjust the vision
  • Senior people at research startups are those who can hold both timescales simultaneously without collapsing one into the other

Insights

The insight that eval gains and feature shipments are not interchangeable backlog items — because one may require methodological pivots while the other requires sprint planning — explains a common startup failure: treating research milestones as product milestones and then being surprised when the team feels like it is constantly failing. The two-timescale framework is directly applicable to ML research teams running ablation studies (sprint = engineering) while pursuing a novel architecture (vision = research).

Connections

Raw Excerpt

“Achieving an x% gain on an eval is not the same kind of milestone as shipping a customer feature request in the UI. The eval gain might require a methodological pivot and weeks of staring at model outputs.”

Graph View