North Star Exploration
Turn aimless tech-tinkering into learning by chasing an arbitrary-but-real goal and being stubborn about reaching it.
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 3
- Confidence
- 90%
Schillace's antidote to unproductive 'just play with it' advice. Pure poking at a new technology rarely produces anything valuable, but a genuine goal — even a goofy or arbitrary one — forces you into the hard edges where real learning happens. The discipline is picking a concrete target that leads somewhere at least mildly interesting, then being a little stubborn about grinding toward it to discover why it's hard.
Origin
Sam Schillace's practice with his Microsoft AI teams (multi-agent systems) and his framing of how to learn any new technology, including generative AI.
Core principles
- 01Just playing with tech without a goal doesn't tend to produce anything super valuable.
- 02An arbitrary goal is fine as long as it's a real goal you're actually trying to reach.
- 03The goal doesn't have to be useful in itself — building something that already exists is fine if the capability you prove out is valuable.
- 04Pick goals that lead somewhere at least a little interesting, in a domain with few distracting side-issues.
- 05Stubbornness toward a target surfaces the surprising results and the real reasons things are hard.
How to run it
- 1
Pick a concrete, arbitrary-but-real target
Choose a specific thing to build or accomplish with the new technology, even if it's goofy. The point is that it's a real goal you'll genuinely try to hit, not open-ended messing around.
Pro tip For learning AI: 'I'm going to build a GPT that can do this part of my job, let's see how close I can get' beats 'I'll mess around with ChatGPT for an hour'.
Watch out 'Play with it for an hour' is not a goal — it won't push you into the hard edges where learning lives.
- 2
Choose a clean domain
Prefer a target in a domain with few incidental distractions so you're testing the capability itself, not fighting unrelated problems.
Pro tip To test how independently multi-agent systems can work, Schillace uses programming: 'you just give something a python environment and a file system and that's it' — no connectivity issues to muddy the signal.
- 3
Grind stubbornly and observe why it's hard
Push toward the goal for a bounded stretch (e.g. a week), staying attentive to surprising failures and the specific reasons progress stalls. The friction is the lesson.
Pro tip Pick the enabling technology partly by popularity if your aim is to learn — lots of other people knowing it means more help and momentum.
Watch out There's no single magic bullet; expect to read a lot, think a lot, and choose your battles.
In the wild
Schillace set his multi-agent programming system the goal of writing vi in Python — a task an intern could do over a summer. Vi already exists, so the artifact is useless, but a system that can self-monitor, self-correct, and build something that complex would be genuinely valuable.
→ The pursuit produced real insights and spun off useful tools — a debugger agent that explains what broke, and an agent that documented the codebase reasonably well.
Chasing better multi-agent cooperation, the team gave agents a shared whiteboard as working memory 'just like a person' to see if they'd cooperate better.
→ It made the agents noticeably smarter — a surprising result they'd have missed in a pessimistic, goal-less frame of mind.
Common mistakes
Goal-less playing
Poking at a technology with no target rarely yields anything valuable; you never reach the hard edges that teach you.
Picking a goal that leads nowhere
An arbitrary goal is fine, but choosing one with no interesting implications wastes the effort — aim at something that at least mildly matters in your domain.
Is it for you?
Best for
Engineers, PMs, and builders trying to learn a fast-moving new technology (like generative AI) and find where its real opportunities are.
Not ideal for
Situations demanding immediate shippable output on a fixed roadmap where exploratory grinding isn't affordable.
From the transcript
“I pick what I call North stars that I think are like interesting useful things to get to rather than just messing around”
“so is that useful by itself no because we already have VI doesn't matter but like if we build a system of programming agents that…”
“just playing you know just playing with tech without being focused doesn't tend to produce anything that's super valuable but picking these kind of even…”
“we gave them whiteboard working memory like as a shared working memory thing to fix this problem and that turns out to make them much…”
“picking in some arbitrary goal and being a little bit stubborn about trying to get yourself to it is a good way to learn stuff”
From the episode
How to be more innovative
Sam Schillace (Microsoft deputy CTO, creator of Google Docs)