The belief
Capability is not the same thing as size.
The prevailing story of AI is a story of scale: bigger models, more data centers, more money — and a shrinking circle of organizations that can afford to take part.
I started Inverse Scaled Intelligence from the opposite conviction. The most important intelligence is the kind ordinary people, students and small teams can actually run.
Scale has become a proxy for progress, and cost has become a moat. I think both assumptions deserve to be challenged. If a small model can do real work — and show that work plainly — then intelligence doesn’t have to be something you rent from a handful of giants.
It can be something everyone owns.
Where it comes from
It started with a question the industry mostly stopped asking: how much of all that scale is actually necessary?
Every year the frontier gets bigger and more expensive, and every year the circle of people who can build with it — or even afford to use it freely — gets smaller.
So I set out to test the opposite bet: that careful design, honest measurement and relentless attention to cost could deliver real, checkable intelligence without a trillion-dollar balance sheet behind it.
Inverse Scaled Intelligence is that bet, made in public.
How I work
Seven habits, kept on purpose.
- Measure before believing.
- I check every result. I compare on equal footing, never on flattering terms, and I run every evaluation in full instead of cherry-picking.
- Publish the misses.
- Wrong answers go right next to right ones. A model you can’t see fail is a model you can’t trust.
- Spend like it matters.
- I treat compute like real money: the cheapest hardware that does the job, nothing left idling, every hour accounted for. Affordability isn’t only a goal for the people using it; it’s a discipline in how I build.
- Build for the whole range.
- The same small model works through market mathematics, logic and code, chess and puzzles, and the sciences — because useful intelligence shouldn’t be confined to one narrow trick.
- Question every default.
- The standard answer is usually the expensive one. I make each piece justify its cost, and anything that doesn’t earn its place goes.
- Let the evidence win.
- I test hunches instead of defending them. When an experiment says I was wrong, the idea goes — and I write the finding down, so I never pay for the same mistake twice.
- Protect the progress.
- I don’t attempt anything risky until the work is safely backed up and verified. Progress is earned slowly, so I never gamble it.
What “usable by all” means to me
Four words, taken literally.
- Affordable
- A cost per answer that rounds to nothing, so nobody has to ration curiosity.
- Accessible
- Runs on everyday hardware, not a data center you’ll never see.
- Honest
- Every answer shown as it is, right or wrong, with the time it took — so trust is earned, not assumed.
- Open to everyone
- The student in a library, the researcher without a grant, the two-person startup, the school on the other side of the world.
I want intelligence to be a tool anyone can build with, not a luxury rationed by a few.