Experiments settle it.
A design model is only as good as its feedback. We connect ours to real atom-scale instruments early, so measurements decide what works, not citations.
Xymos
Self-replicating nanites.
Xymos is building AI that designs and runs fabricators: machines that work with individual atoms, and build more of themselves. It's the most direct route we know from intelligence to real physical capability.
01 Goal
AI is getting smarter faster than it is getting reach.
AI systems are becoming very good at reasoning, code and analysis. What they can change in the physical world is still limited by human-scale tools: labs, factories, robots, and the people who run them. Experiments are slow, expensive and mostly manual.
We think reach, more than intelligence, is now the ceiling on what AI does for the world.
We think the most general interface to the physical world is the atomic scale, because everything is built from atoms. A system that can place and measure them directly, and build better tools for doing so, is not limited to any one application. It could apply to medicine, materials, energy and security.
The thesis
The end goal is not nanofabrication. It is an AI that can act on the physical world directly, at the smallest scale matter allows.
02 Approach
Nanofabrication is how.
As far as anyone knows, physics permits atomically precise manufacturing. The idea is more than sixty years old, and researchers have been placing single atoms by hand since 1990, but it has never become general-purpose manufacturing. We think the blocker is design and search, which is what a very capable AI is for.
- Step 1
Build the designer.
Frontier models are the springboard, where design and search start. What runs closest to the atoms has to be small and specialized, because a nanoscale machine has almost no room for computation.
Open problem - Step 2
Make it controllable.
Aligning a general AI is unsolved. Controlling a replicator is mostly engineering: what it can build, what it can feed on, how far a lineage can go. We'll design that in from the start rather than bolt it on.
Open problem - Step 3
Give it reach.
Connect it to instruments that place and measure single atoms, so it can run experiments on the physics directly instead of reasoning about it from papers.
Not started - Step 4
Close the loop.
Once it can fabricate at the nanoscale, each round of fabrication should improve the next, including copies of itself. Capability then scales by replication, not by building factories.
Not started
03 Tractability
A closed world, not an open road.
A self-driving car has to work in an open world: weather, roadworks, other drivers, people doing something new every day. That long tail is most of the engineering, and getting through it has taken years of work at enormous scale.
A self-replicating fabricator works in a closed one. The chamber is ours, the feedstock is chosen, and the physics doesn't change. The design problem is hard but bounded, much of it can be searched in simulation before anything is built, and once one works, replication does the scaling.
The uncomfortable corollary: an uncontrolled replicator isn't a decades-long project either. Building one with control designed in should cost only somewhat more. We think it should be built that way, first, on purpose.
04 Principles
How we'll work.
Limits are physical.
Every replicator we design has hard limits on what it can build, what it can consume and how many generations it can run. They are properties of the hardware, not settings in software.
Small models near the machine.
Large models help us search the design space. Whatever runs on or beside a machine is small, specialized and simple enough to inspect.
Show the work.
We'll publish designs, measurements and failures as we go, so anything claimed on this page can be checked.
05 Contact
Contact.
If you work on alignment, nanofabrication or frontier AI, or want to support this, we'd like to hear from you.
hello@xymos.org