encepticWe’re building toward an architecture that spends compute where it counts, keeps what it learns, and reuses verified discoveries, so capability can grow without the bill growing with it.
One intelligence, different kinds of effort
Some questions want a fast answer. Unfamiliar ones might need a search, an experiment, or a second check. We’re working on systems that pick the right amount of effort for the question and know when to stop.
Experience that remains useful
Picture a small desk backed by a huge archive. That’s the memory we’re after: pull up the relevant page with its sources attached, instead of rereading the whole library every time.
Discovery that pays forward
Costly thinking should pay rent. When a hard problem gets solved and checked, we want to pack the method into a reusable procedure, its limits written on the box, with a fallback for when it doesn’t fit.
Where the evidence stands
In narrow lab tests on synthetic tasks, choosing what to compute and when to quit has worked encouragingly well. That is all it shows. No general reasoning ability, no finished architecture, no proven real-world savings.
Lasting memory, reusable discovery, and broader adaptability are still open questions. The architecture isn’t settled. There is no release date and no pricing.
What we intend to share
When results come, we’ll publish the method, the baselines, what it cost to run, and where it falls short. Bigger claims will need repeatable tests on unfamiliar tasks, with the full bill attached.
About the opening film
The film opens on scaled dot-product attention and the 2017 paper Attention Is All You Need, then follows how far one idea traveled. Our own goals are still ahead of us.