The Decisive No Filter Framework | Generated by AI
Question: How do you make decisive choices about what to focus on and collaborate on — when many things seem interesting and useful, and many people want your help? How did Steve Jobs cut at Apple? How do you develop a clear “no” filter?
Answer:
The Core Problem
You’re technically versatile, AI-native, and visibly capable. That means everyone wants a slice — collaborations, side projects, consulting asks, “quick favors.” The trap is that each individual thing looks good. The cost isn’t any single yes. It’s the aggregate of yeses destroying your compounding trajectory.
What Jobs Actually Did
When Jobs returned to Apple in 1997, Apple had 350 products. He cut to 10. The mental model he used was brutally simple:
“Deciding what not to do is as important as deciding what to do.”
The actual mechanism:
- He drew a 2x2: Consumer / Pro × Desktop / Portable. Four cells. One product per cell. Everything else died.
- The filter wasn’t “is this good?” — it was “does this fit the 4 boxes?” If not, no, regardless of merit.
The insight: a pre-committed framework kills individual negotiation. You don’t decide per-case. You decide the framework once, and the framework decides for you.
Build Your Own 2x2
For you specifically, given your trajectory toward AI engineering, something like:
| Trains my AI depth | Doesn’t | |
|---|---|---|
| Ships something real | ✅ DO IT | ⚠️ only if very fast |
| Doesn’t ship | ⚠️ only if deep learning | ❌ say no |
Everything that lands in the bottom-right is an automatic no. This removes the emotional cost of deciding — you just point to the framework.
The “Interesting and Useful” Trap
Things can be both interesting and useful and still be wrong for you right now. The question isn’t quality of the thing — it’s opportunity cost.
You have one real constraint: focused time on the trajectory toward AI engineering depth. Every collaboration that doesn’t compound that trajectory is trading your future for someone else’s present.
Practical heuristic:
“Will this make me a better AI engineer in 6 months, or will I just have helped someone else ship faster?”
If the answer is the latter — even if it pays — it’s a tax on your trajectory.
How to Know What You Actually Want
Jobs had a trick for this: the “Hell Yes or No” rule (popularized by Derek Sivers, practiced by Jobs-era Apple culturally).
If your gut reaction to a request isn’t immediately “hell yes” — it’s a no. The reasoning:
- Your real priorities generate immediate pull. You don’t need to convince yourself.
- Ambivalence is a signal that it doesn’t fit your core, not that you need more information.
For you: when someone asks for a collaboration, notice your first 2 seconds. Genuine excitement about the AI/technical depth involved? Yes. Mild interest + guilt about saying no? No.
How to Actually Say No
Most people struggle with no because they make it personal. Make it structural:
Template:
“I’m heads-down on [specific thing] for the next [time period] and I’ve committed to not taking on new collaborations until then. I have to pass — but [optional genuine redirect].”
You’re not rejecting them. You’re pointing to a pre-existing constraint. This is honest (you should actually have that commitment), and it’s socially clean.
For async asks (GitHub issues, DMs, email):
“Not taking this on — outside my current focus area.”
One line. No elaboration. Elaboration invites negotiation.
The Karpathy Model (Relevant to You)
Karpathy runs a near-zero-collaboration model for deep work. He doesn’t co-author papers, rarely does consulting, ignores most inbound. His output — nanoGPT, LLM.c, the video essays — is maximally high-leverage because his focus is undivided.
The lesson isn’t “be a hermit.” It’s: your most important work requires uninterrupted compounding. Each context switch doesn’t just cost the switch — it resets the depth accumulation.
For you: training a model from scratch, building a real agent system, deeply understanding DeepSeek MoE internals — these require 4-6 hour unbroken blocks, sustained over weeks. Every collaboration that fragments that is a direct tax on the work that matters.
Practical System
- Define your 3 current focus areas explicitly. Write them down. Everything incoming gets evaluated against these 3 only.
- Default is no. Yes requires an active override — not the reverse.
- Time-box reviews. Once a month, reassess whether your 3 focus areas are still right. Don’t reassess per-request.
- Create a “someday” list. When something interesting comes in that doesn’t fit, write it there. This lets you say no without loss anxiety — it’s not gone, just deferred. Most things on that list you’ll never do, and that’s fine.
- Protect your mornings. First 4 hours = your focus areas only. Collaboration, communication, requests = afternoons.
For You Specifically
Your actual 3 right now are probably something like:
- AI engineering depth — transformer internals, training, fine-tuning, MoE
- Shipping real AI systems — agents, LLM pipelines, tools that do real work
- Public knowledge compounding — blog, notes, building reputation as an AI engineer
Anything incoming gets asked: does this directly serve one of these three? If not — no. Even if interesting. Even if useful. Even if the person is great.
The goal isn’t to do fewer things. It’s to do the right things with enough depth that they actually compound into the career you’re building.