Northstar Problem Commitment
Pick one foundational building-block problem and commit years to it, not a scatter of trendy ones.
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 83%
A research and career strategy: rather than chasing many problems, identify the single foundational problem that is a building block for everything downstream, and commit to it as your northstar for years. Use it when choosing a research agenda, a company's core bet, or a long-term area of mastery.
Origin
Early in her career Li chose object recognition as her northstar because objects are the level at which humans 'interpret, reason and interact' with the world — making it a building block for the whole perceptual world. That single commitment led to ImageNet and, indirectly, modern deep learning.
Core principles
- 01Choose a problem that is a building block others must stand on, not a leaf feature.
- 02Anchor the choice in how the real system (humans, the world) actually operates.
- 03Commit for the long haul; northstar problems reward years, not sprints.
- 04The right foundational problem generates unexpected downstream breakthroughs.
How to run it
- 1
Find the building-block layer
Ask what level everything else depends on. For perception it was objects, because we interact with the world 'more or less at the object level', not the molecular level.
Pro tip Test candidates by asking: if this were solved, how many other problems get easier?
- 2
Declare it your northstar
Commit publicly and internally to one problem as the organizing goal of your work.
Watch out Resist diluting focus across many fashionable problems at once.
- 3
Interrogate the missing ingredient
Ask what critically overlooked input is blocking progress — for Li, it was big data, which the field had ignored.
Pro tip Study how nature/humans solve it; human learning and evolution are 'big data' processes, which pointed her to data.
- 4
Build the enabling asset and share it
Create the resource that unblocks the whole field, then open it up to compound the effort.
Pro tip Open-sourcing ImageNet and running an annual challenge turned one lab's bet into a field-wide movement.
In the wild
Li identified object recognition as a northstar building block, diagnosed missing big data as the blocker, curated 15M labeled images across 22,000 concepts, and open-sourced it with an annual challenge.
→ The dataset became the substrate for the 2012 deep-learning breakthrough and the current approach to scaling AI.
Common mistakes
Optimizing the model instead of the foundation
For years the field poured effort into ever-fancier models while ignoring the foundational missing ingredient (data). Focus on the building block, not just the tool.
Is it for you?
Best for
Researchers, founders, and specialists choosing a decade-defining area to commit to.
Not ideal for
Short-horizon or exploratory work where breadth and fast pivots matter more than deep commitment.
From the transcript
“my students and and I are very committed to a northstar problem which is solving the problem of object recognition because it's a building block…”
“a very critically overlooked ingredient of bringing AI to life is big data”
From the episode
The Godmother of AI on jobs, robots, and why world models are next
Dr. Fei-Fei Li